80 GB of AI memory in a pocket-sized device sounds compelling. But what can you actually do with it? Over the past weeks, I experimented with two @TiinyAILab #AI devices prosumer.one/guides/tiiny-po… alongside a small #localAI lab running @AMD and @NVIDIA GPUs, @Linux\@Olares_OS and @AnythingLLM.
⟴ The most interesting result was on the #Tiiny White. I successfully ran GPT-OSS-120B locally. During one recorded run, the device reported:
⇗ NPU memory: 46.1 / 47.9 GiB
⇗ System RAM: 20.5 / 31.1 GiB
⇗ One practical planning response: approximately 38 seconds
These are device-level memory readings and one request's elapsed time — not isolated model-memory measurements or a standardized inference benchmark.
⟴ On the Tiiny Black, @Alibaba_Qwen 3.6-35B-A3B also ran successfully via its local API. One diagnostic request returned 282 completion tokens with a total elapsed time of approximately 47.9 seconds.
My separate @AMDRadeon AI PRO #R9700 system achieved nitter.gq/NansenID/status/212272… approximately 27 tokens/second generating with Qwen2.5-32B Q4_K_M in earlier testing. That number isn't directly comparable: different models, runtimes, quantization and workloads.
The architectural distinction was more interesting than the raw numbers: Tiiny exposes separate NPU and system-memory pools, while my conventional GPU systems have different memory limits and software flexibility.
⬤ I also managed to connect both devices to a shared AnythingLLM environment and verify stored responses. But the next level — useful autonomous workflows, independent AI collaboration and meaningful memory reuse — remained unproven. A later attempt ended with HTTP 502 responses from both devices.
Haven't tried yet: #Odysseus nitter.gq/NansenID/status/208142… etc = sooner or later \\
⮆ What I learned: fitting a 120B model into a compact local device is an impressive engineering accomplishment (thnx, Tiiny-dreamteam)! Turning that capability into dependable, useful work is an entirely different challenge...
That is the benchmark I care about most: prosumer.one/guides/agent-fi…
These are observations from my own setup, not a controlled hardware review or a conclusion about everyone else's experience.
AI-friendly → agent-friendly (agentic-first?) prosumer.one/guides/agent-fi… One computer. Two operators. And apparently no one wrote the contract.
Operator #1: me. I decide why the machine exists, what it is allowed to do, and when to pull the plug.
Operator #2: my AI assistant. It can read logs, trace failures and help repair a broken GPU setup. That's already useful. But “it can access my computer” is NOT the same as “it is authorised to do this particular thing.”
Our #DGXSpark experiment demonstrated assisted recovery. It did not demonstrate safe delegation to another human, scoped agent access, or reliable revocation. Those are different questions, and I would rather test them than bury them under a glossy “autonomous agent” slide.
The hardware industry likes its TOPS, tokens/sec and RAM charts. Fine. Here's my less fashionable metric: how much of the owner's life does the machine give back?
An agent with root and a good demo is not a product strategy. It's a future incident report with nicer typography.
I want a second operator with an actual mandate: who authorised it, for what, within which limits, until when — and how to take that authority back. Not just an agent that *can*, but an agent that *may*.
That's where personal AI, the Prosumer idea and AuthorityClaw start to meet: solodkiy.cv/agent-ready-comp… Personal isn't merely where the model runs. It's who remains in control.
The more interesting part began when I installed @OlaresOS on the same #DGXSpark machine (instead of native OS). The installation, the repair and the evidence:
solodkiy.cv/dgx-spark-olares… Now the question wasn't “Can it run a model?” It was “Can I — with the AI assistant I already use — make this thing work for me without becoming its full-time sysadmin?”
Drivers. Recovery. Identity. Remote access. The gap between a working GPU and a working *experience* is where the real product lives.
14%. That's where my “AI-first” computer decided to turn itself into a Linux certification exam.
Secure Boot rejected the NVIDIA module → the driver installation failed → NVML was missing → Olares couldn't confirm GPU readiness. The installation stopped. The user-facing option was essentially “Uninstall and try again.”
Very reassuring. Almost a spa experience.
I handled UEFI. My AI operator traced the failure chain, repaired the @nvidia software stack and verified CUDA/NVML. We reset the *partial Olares installation*, not the entire machine, and eventually got through activation.
Then, finally, actual work: Market-owned @Ollama ran Qwen inference; core ComfyUI generated an image on GB10. Optional plugins failed, and one inspected FreeToken package was amd64-only. An honest field test: useful results, visible limits.
This is why I find “AI-friendly” a dangerously cheap label. Don't tell me your platform has an AI button. Tell me whether it can identify *which dependency broke*, explain *what a reset destroys*, and let an assistant fix what is fixable without pretending that my consent is a decorative checkbox.
I don't need frictionless marketing. I need friction I can understand — and a route back to doing something useful.
47
The more interesting part began when I installed @OlaresOS on the same #DGXSpark machine (instead of native OS). The installation, the repair and the evidence:
solodkiy.cv/dgx-spark-olares… Now the question wasn't “Can it run a model?” It was “Can I — with the AI assistant I already use — make this thing work for me without becoming its full-time sysadmin?”
Drivers. Recovery. Identity. Remote access. The gap between a working GPU and a working *experience* is where the real product lives.
14%. That's where my “AI-first” computer decided to turn itself into a Linux certification exam.
Secure Boot rejected the NVIDIA module → the driver installation failed → NVML was missing → Olares couldn't confirm GPU readiness. The installation stopped. The user-facing option was essentially “Uninstall and try again.”
Very reassuring. Almost a spa experience.
I handled UEFI. My AI operator traced the failure chain, repaired the @nvidia software stack and verified CUDA/NVML. We reset the *partial Olares installation*, not the entire machine, and eventually got through activation.
Then, finally, actual work: Market-owned @Ollama ran Qwen inference; core ComfyUI generated an image on GB10. Optional plugins failed, and one inspected FreeToken package was amd64-only. An honest field test: useful results, visible limits.
This is why I find “AI-friendly” a dangerously cheap label. Don't tell me your platform has an AI button. Tell me whether it can identify *which dependency broke*, explain *what a reset destroys*, and let an assistant fix what is fixable without pretending that my consent is a decorative checkbox.
I don't need frictionless marketing. I need friction I can understand — and a route back to doing something useful.
NVIDIA gave me a #DGXSparkFE as a prize. The field story, including my own stupid mistakes:
solodkiy.cv/dgx-spark.html
Naturally, my first experiment was to switch it OFF while trying to switch it ON:) A promising start for the local-AI revolution (my fault). It had powered up when I plugged it in. I pressed the power button because that's what power buttons are for. Apparently not.
Then came the serious stuff: gpt-oss:120b on native DGX OS, plus image, speech and vision workloads. The hardware could compute. Surprise: @NVIDIA makes decent hardware.
#JensenHuang has a habit of making yesterday's ambitious questions look embarrassingly small to me. I don't need to borrow his answers. I need to take that irritation back to my own work and build something.
I like benchmarks. I just refuse to confuse them with a product. @NVIDIAAI @NVIDIARTXSpark
NVIDIA gave me a #DGXSparkFE as a prize. The field story, including my own stupid mistakes:
solodkiy.cv/dgx-spark.html
Naturally, my first experiment was to switch it OFF while trying to switch it ON:) A promising start for the local-AI revolution (my fault). It had powered up when I plugged it in. I pressed the power button because that's what power buttons are for. Apparently not.
Then came the serious stuff: gpt-oss:120b on native DGX OS, plus image, speech and vision workloads. The hardware could compute. Surprise: @NVIDIA makes decent hardware.
#JensenHuang has a habit of making yesterday's ambitious questions look embarrassingly small to me. I don't need to borrow his answers. I need to take that irritation back to my own work and build something.
I like benchmarks. I just refuse to confuse them with a product. @NVIDIAAI @NVIDIARTXSpark
#DGXSparkFE @NVIDIAAI arrived on 22 Oct 🔥 and tested 🎇 works well btw w @OlaresOS too 🚀
Every time I think I need another expensive GPU, someone comes up with a cleverer way to use the hardware I already own!
[1] A 35B MoE on my 12GB #RTX3060 with #FreeToken? Done.
[2] @coldniko's #Strata answering in English and Russian on my #DGXSpark? Also done, although our memory guard ended that particular adventure early.
⟱
[3] And now @zili_meng's @WiCiHQ has released InferSSD: github.com/wici-ai/inferssd ⇗
↳ The idea I like: instead of insisting that all the experts of a huge MoE model live in RAM or VRAM, let the SSD do more of the work.
↳ Stream the experts you need, cache what you can, and see what your existing computer can manage.
⟴ Their own (not mine this time) @NVIDIARTXSpark numbers for an 816B model are around 2 tokens/sec.
⟴ That's not a fast chat. But the interesting part is that this class of model can run at all in that memory budget, right?:)
I haven't tested InferSSD yet. I just genuinely like this direction — and the fact that #WiCi prosumer.watch/dossiers/wici… published the engine openly.
Good work, #WiCi nitter.gq/NansenID/status/210751… ⇉ I hope you keep pushing it ⇮
I'm interested in useful AI, not becoming a museum of GPUs:) I'd be glad to share some messy, real-world user observations once I have something worth reporting. @NVIDIAAI
#Strata worked on my @nvidia #DGXSpark. (Then I shut it down, but it's not important here.)
I've been playing with #localAI on a rather improbable collection of prosumer.one 's mini-PCs and GPUs. Sometimes the hardest part isn't running a model. It's getting all the software — and the AI assistants supposedly helping me — to cooperate!
So I tried @coldniko's Strata github.com/Niko1221/Strata ⇯ on my @NVIDIARTXSpark running (first it's native OS - then) #OlaresOS: nitter.gq/NansenID/status/210601… ↬
It actually answered:) @Alibaba_Qwen 3.8-Flash-Next, about 11.6–15.2 tokens/sec in those two runs!
Then our own memory safety limit stopped the experiment before the next task:)
That's the whole story: two real answers, no long-term service, no disaster:)
On another #RTX3060 machine, I got the Strata engine installed, but didn't have enough free memory to run the large MoE model without disturbing the other things on that PC with nitter.gq/SlavaSolodkiy/status/2… (my bad, not Strata's). Another small machine didn't even pass our resource check - it's ok.
I'm not running a professional GPU lab, and I'm not trying to build a definitive leaderboard.
I'm trying to figure out what I can actually do with hardware I already have.
Then: #FreeToken showed me that a 35B MoE could run on a 12GB GPU nitter.gq/SlavaSolodkiy/status/2… ⇨ Now: Strata gave me another useful piece of the picture!
and...
There's a bigger question here than tokens/sec: ↹ How much real work can this software save me — and how much of my own time does it take to make it useful?
I appreciate the people building Strata and similar engines! These are the projects that keep changing what's possible on ordinary machines.
If the developers are curious about what happened in my slightly chaotic setup, I have the logs:)
@NVIDIAAI
156
#Strata worked on my @nvidia #DGXSpark. (Then I shut it down, but it's not important here.)
I've been playing with #localAI on a rather improbable collection of prosumer.one 's mini-PCs and GPUs. Sometimes the hardest part isn't running a model. It's getting all the software — and the AI assistants supposedly helping me — to cooperate!
So I tried @coldniko's Strata github.com/Niko1221/Strata ⇯ on my @NVIDIARTXSpark running (first it's native OS - then) #OlaresOS: nitter.gq/NansenID/status/210601… ↬
It actually answered:) @Alibaba_Qwen 3.8-Flash-Next, about 11.6–15.2 tokens/sec in those two runs!
Then our own memory safety limit stopped the experiment before the next task:)
That's the whole story: two real answers, no long-term service, no disaster:)
On another #RTX3060 machine, I got the Strata engine installed, but didn't have enough free memory to run the large MoE model without disturbing the other things on that PC with nitter.gq/SlavaSolodkiy/status/2… (my bad, not Strata's). Another small machine didn't even pass our resource check - it's ok.
I'm not running a professional GPU lab, and I'm not trying to build a definitive leaderboard.
I'm trying to figure out what I can actually do with hardware I already have.
Then: #FreeToken showed me that a 35B MoE could run on a 12GB GPU nitter.gq/SlavaSolodkiy/status/2… ⇨ Now: Strata gave me another useful piece of the picture!
and...
There's a bigger question here than tokens/sec: ↹ How much real work can this software save me — and how much of my own time does it take to make it useful?
I appreciate the people building Strata and similar engines! These are the projects that keep changing what's possible on ordinary machines.
If the developers are curious about what happened in my slightly chaotic setup, I have the logs:)
@NVIDIAAI
#DGXSparkFE @NVIDIAAI arrived on 22 Oct 🔥 and tested 🎇 works well btw w @OlaresOS too 🚀
I've run a 120-billion-parameter #AI model on a 24GB GPU. And I've seen a 12GB GPU run a 35B model surprisingly well.
⇾ Meanwhile, I've also spent hours fighting software that couldn't make proper use of perfectly capable hardware.
⇾ The more local AI systems I try, the less I believe that GPU specifications tell the whole story.
Over the past few months, my somewhat chaotic collection of hardware has included:
⇾ @AMDRadeon cards with 16GB and 24GB - and now ↯ 32GB of VRAM too!
⇾ @Intel Arc B580 with 12GB.
⇾ Two @NVIDIA #RTX3060 with 12GB each.
⇾ An #OlaresOne with a ↯ 24GB @NVIDIAAI #RTX5090 Laptop\Mobile GPU.
⇾ An @NVIDIARTXSpark #DGXSpark with ↯ 128GB of unified memory.
↯↯ And two tiny, self-contained local-AI-first rigs called @TiinyAILab with 80GB of unified LPDDR5X memory (plus, a 12-core ARMv9.2 processor, custom NPU, and a 1TB SSD in a roughly 300-gram chassis with only 30W (!) consumption).
⇶ Some of my favourite discoveries:
1. A 12GB #RTX3060 ran a 35B MoE model.
⟿ #FreeToken + @Alibaba_Qwen 3.6-35B-A3B, around 20.5 tokens/sec decode, using only about 7GB of VRAM.
⇵ Not because the entire model magically fitted into memory. Because the runtime made intelligent use of CPU, RAM and GPU execution.
2. A 24GB NVIDIA GPU ran GPT-OSS-120B.
⟿ On my #OlaresOne, FreeToken delivered around 54.4 tokens/sec of server-side decode.
3. The 128GB #DGXSpark ran the same 120B model through @ollama.
⟿ Around 43.6 tokens/sec on warmed requests under its original DGX OS.
⇵ Different runtime and execution strategy — not a fair GPU speed comparison, but a fascinating demonstration of how much the software stack matters.
4. An @Intel @arc #B580 did more than just display graphics:)
⟿ I ran an @OpenVinoAI reranker, then #Nemotron 3 Nano 4B through #Ollama/#Vulkan, entirely on the GPU. It even completed a real tool-calling round trip.
⇵ Getting everything integrated with Olares Router was another story.
5. My latest 32GB @AMD #R9700 runs a dense 32B model entirely on the GPU.
⟿ About 27 tokens/sec for @qwenmax 2.5-32B Q4_K_M.
⇵ And it now serves a separate 27B model to my local @AnythingLLM setup.
None of these numbers form a scientifically controlled GPU leaderboard. They come from different experiments, models and inference engines.
But they're real experiments on real hardware, including the failures.
And they've changed the questions I ask.
Not simply: How many GB of VRAM does this machine have? But:
What can I actually run? How much memory can the runtime exploit? Can the model use tools? Can I connect it to my other machines? Does it stay working after the benchmark?
And, increasingly: How much of my own time does it consume before becoming useful?:)
That last metric might be the most underrated benchmark in AI, hahah.
Especially when the AI agents themselves are supposed to save us time, right?
#LocalAI #OpenSourceAI #Inference #AMD #NVIDIA @QwenDevs
I used to run two @AMD GPUs already at home: #RX7800 XT 16GB → #RX7900 XTX 24GB → A few days ago, I installed a @PowerColor @AMDRadeon #AI PRO #R9700 with 32GB of VRAM.
The hardest part wasn't getting the GPU to work. It was getting my AI agents to figure out that it was already working. My setup is slightly ridiculous:
→ @Hi_MINISFORUM's @AMDRyzen mini-PC and their OCuLink DEG1 eGPU dock (with @GlobalLianli #LianLi 1200W PSU)
→ @OlaresOS / @Linux / @kubernetesio
A growing collection of #AI assistants supposedly capable of operating all of this:)
The hardware booted. The GPU was detected. The driver worked. And then the fun began.
One AI assistant couldn't access the Olares terminal. Another tried a different route and hit a 404. Another ran out of usage credits before completing the diagnosis.
At some point, I realised I wasn't benchmarking the GPU anymore.
I was benchmarking the AI agents trying to benchmark the GPU:)))
Eventually, @claudeai's coding environment found a working path through #Kubernetes.
And suddenly everything worked! O_o
@Alibaba_Qwen 2.5, Q4_K_M, running locally on the R9700:
→ 7B — 102.3 tokens/sec
→ 14B — 54.3 tokens/sec
→ 32B — 27.1 tokens/sec
Three runs per @qwenmax model. All layers on the GPU. No crashes or out-of-memory errors. Maximum junction temperature: 69°C.
→ The 32B model used around 21GB of VRAM, leaving roughly 11GB available.
→ All through an OCuLink connection running at PCIe 4.0 ×4.
And this wasn't just an isolated benchmark. The R9700 is now running a separate 27B model as the fourth backend in my local @AnythingLLM setup, alongside two @nvidia #RTX3060 machines and a dedicated @TiinyAILab devices.
My takeaway?
→ The #AMD hardware was less troublesome than I expected:)
→ The AI orchestration was more troublesome than it had any right to be:(
And the most useful breakthrough wasn't a new model or driver. It was finding an AI execution environment that could actually operate the machine.
Next experiment: a used AMD Instinct MI50 with 32GB of HBM2... Can an older accelerator deliver enough of the same practical utility (at less than a third of the price)?
That's a comparison I'm genuinely curious about.
#LocalAI #AMD #ROCm #AIInfrastructure @ClaudeDevs
4
2
101
I used to run two @AMD GPUs already at home: #RX7800 XT 16GB → #RX7900 XTX 24GB → A few days ago, I installed a @PowerColor @AMDRadeon #AI PRO #R9700 with 32GB of VRAM.
The hardest part wasn't getting the GPU to work. It was getting my AI agents to figure out that it was already working. My setup is slightly ridiculous:
→ @Hi_MINISFORUM's @AMDRyzen mini-PC and their OCuLink DEG1 eGPU dock (with @GlobalLianli #LianLi 1200W PSU)
→ @OlaresOS / @Linux / @kubernetesio
A growing collection of #AI assistants supposedly capable of operating all of this:)
The hardware booted. The GPU was detected. The driver worked. And then the fun began.
One AI assistant couldn't access the Olares terminal. Another tried a different route and hit a 404. Another ran out of usage credits before completing the diagnosis.
At some point, I realised I wasn't benchmarking the GPU anymore.
I was benchmarking the AI agents trying to benchmark the GPU:)))
Eventually, @claudeai's coding environment found a working path through #Kubernetes.
And suddenly everything worked! O_o
@Alibaba_Qwen 2.5, Q4_K_M, running locally on the R9700:
→ 7B — 102.3 tokens/sec
→ 14B — 54.3 tokens/sec
→ 32B — 27.1 tokens/sec
Three runs per @qwenmax model. All layers on the GPU. No crashes or out-of-memory errors. Maximum junction temperature: 69°C.
→ The 32B model used around 21GB of VRAM, leaving roughly 11GB available.
→ All through an OCuLink connection running at PCIe 4.0 ×4.
And this wasn't just an isolated benchmark. The R9700 is now running a separate 27B model as the fourth backend in my local @AnythingLLM setup, alongside two @nvidia #RTX3060 machines and a dedicated @TiinyAILab devices.
My takeaway?
→ The #AMD hardware was less troublesome than I expected:)
→ The AI orchestration was more troublesome than it had any right to be:(
And the most useful breakthrough wasn't a new model or driver. It was finding an AI execution environment that could actually operate the machine.
Next experiment: a used AMD Instinct MI50 with 32GB of HBM2... Can an older accelerator deliver enough of the same practical utility (at less than a third of the price)?
That's a comparison I'm genuinely curious about.
#LocalAI #AMD #ROCm #AIInfrastructure @ClaudeDevs
The tool is no longer the bottleneck. The question might be.
Give me an AI that answers my question in half the time: useful.
Give me an AI that helps me notice I was asking the wrong question: now we might be getting somewhere.
Three more thoughts from hearing #JensenHuang (suddenly not yesterday, in Berlin) in London on 14 October. Recollections and interpretation, not a transcript. The Bach analogy is entirely my fault.
1. DON'T JUST USE THE MACHINE. BUILD THE MACHINE.
The idea that stayed with me wasn't simply producing more things with AI. It was being able to create your own means of producing them.
Using software to do your work is one thing. Building the software through which your work gets done is another.
One makes you a faster operator. The other lets you question the operation.
Imagine wanting to play Bach, but first having to prove you can read every mark in the score. Now imagine that technical entry barrier becoming less decisive.
Not “engineering no longer matters”. More: “technical notation is no longer the only admission ticket.” You still need judgment. What do you want to play? Can you tell whether it sounds right?
2. THE AGE OF AGENTS MAY ARRIVE AS A HABIT.
I remember Jensen comparing agents with the early browser era. I took away a psychological point: people may adopt useful agents before the technology becomes immaculate. Repetition turns something unsettling into something ordinary.
Eventually, some of it becomes Tuesday.
My qualification: familiarity is not safety. But asking people to remain permanently frightened isn't a product strategy either.
The system has to earn the right to become boring.
That's why I care about controls that remain in force after the user relaxes. NVIDIA's OpenShell announcement in Berlin yesterday adds a concrete engineering counterpart: containment, monitoring and policy enforcement outside the agent's own discretion.
3. FASTER ANSWERS ARE NOT THE ONLY KIND OF PROGRESS.
“How do I make this payment faster?” becomes “Does the agent still have authority to make it?”
“Where do we find power for a data centre?” becomes “What else could an existing energy site provide?”
Those are different design problems, not more eloquent answers to the old ones.
We already discussed the move from creators and coders to industry-specific agents. The next question is what industry experts choose to build once building becomes easier.
I don't just want AI to agree with my question more efficiently. I want it to help me discover a better one — and build something small enough to test it.
My entirely unsolicited two-word slogan for the @nvidia:
Keep asking.
What would you build if the technical entry barrier were lower? What question would make it worth building at all? @NVIDIAAI
One less technical memory from the evening: Jensen was asking people near the stage whether @nvidia had invested in them. When he got to me, I said “Oops, no!”
His reply, as I remember: “Then I will!” I’m filing that under excellent stage improvisation, not a funding announcement
I was looking forward to #PineaPi by @OpenBMB aka @MiniCPM.
Not just as another mini AI box prosumer.one/ but as a personal node from the people who’ve been serious about efficient on-device models (MiniCPM). That combination is still uncommon.
They’ve paused the hardware project — said they couldn’t hit the experience and price point they wanted, and preferred not to ship a compromise:(
The idea was solid. Wishing the team well as they keep building on the model and tools side!!!
I was looking forward to it. The combination of capable edge models (MiniCPM) and dedicated hardware built around them is still uncommon. Most boxes in this space are either underpowered or overpriced compromises. This one looked like it was trying to avoid that trap.
They said the experience and specs they wanted couldn’t be sustained at a consumer price point, and they chose not to ship something they weren’t proud of. That’s a difficult call. It’s also an honest one. Better to stop than to release a product that would have disappointed the people who were waiting for it. No drama there.
@openbmb_model is continuing with the open-source model and tools work! The underlying capability doesn’t disappear just because the hardware project paused. Good ideas in local and edge AI are still rare enough that it’s worth noting when one of them is tried.
Wishing the team well on whatever comes next. The intention was solid.
54
Two tiny KVMs. A much bigger question: can AI make personal computers genuinely self-manageable? A follow-up to my @SipeedLab NanoKVMs Go & Go+ experiment.
I've connected two physical @OlaresOS-based mini-PCs — a @Hi_MINISFORUM & an @Acemagic_MiniPC — to NanoKVM devices running their own @Tailscale connections.
From my @surface (and @softwincn GPD Pocket 4), outside my home, I can now see both physical @Linux consoles. No dependency on the hosts' operating systems, remote desktops or SSH.
That's already useful. But it's only the beginning.
Here is what I'd like to test next:
1. Can an AI agent recover a computer that has stopped cooperating?
Not another agent clicking around a perfectly functioning desktop.
Imagine: a local AI box becomes unreachable. An agent running elsewhere connects through the hardware KVM, inspects the actual screen, understands what failed, proposes a remedy, requests approval, executes it and verifies recovery.
A controlled, reproducible experiment — not deliberately destroying a working system.
2. Can hardware-level computer use become reliable enough for agents?
NanoKVM already exposes screen capture and keyboard/mouse functions through MCP.
Screen capture and keyboard worked in my initial tests. Mouse control still has USB HID timeouts.
I'd like to investigate reliability across BIOS, boot screens, Linux terminals and Windows — including latency, reconnection and recovery after failed actions.
An agent having eyes and hands is interesting. An agent having reliable eyes and hands is useful.
3. Could Go+ Memory Fabric become a computer's operational memory?
Imagine a device remembering what appeared on its screen before something failed.
Could a local model reconstruct the sequence, identify recurring errors and explain what changed?
I'd particularly like to explore local inference, selective recording, retention controls and privacy. Operational memory shouldn't mean permanent surveillance.
4. What does this mean for prosumers?
This is the question I find most interesting.
I have several mini-PCs running local AI workloads. Most people interested in personal AI don't want to become part-time Linux administrators.
What if a £/$-accessible hardware accessory, a secure network and an AI assistant could eliminate much of the anxiety of running your own infrastructure?
No permanent monitor. No emergency trip home. No mysterious black box requiring expert intervention whenever something goes wrong.
I'd like to measure actual setup friction, cost, reliability and how much technical knowledge remains necessary.
5. And who authorises the agent?
This connects directly to my AuthorityClaw solodkiy.cv/authorityclaw.ht… work.
Seeing a screen is one capability. Changing BIOS settings, entering credentials, rebooting machines or accessing private information are very different capabilities.
The question isn't merely can the agent do it?
It's who authorised it, for which machine, for what purpose, under what limits, and with what evidence afterward?
I'd like to test an explicit permission and audit layer between an AI agent and the NanoKVM MCP interface.
The longer-term idea is simple:
Personal AI infrastructure should be personal not only because you own the hardware, but because you can understand, control and recover it without becoming its full-time administrator.
Two NanoKVMs don't solve that problem. Remote video access is verified; autonomous recovery is not. Separate power management may also be necessary.
But they provide an unusually interesting starting point.
For engineers: an experiment in out-of-band agentic computer control.
For product builders: nitter.gq/SlavaSolodkiy/status/2… a question about making self-hosted AI genuinely usable.
For prosumers: slavasolodkiy.substack.com/p… perhaps one small step toward owning your compute without inheriting a second job.
#LocalAI #NanoKVM #OlaresOS #AgenticAI #Prosumer
2 tiny @SipeedLab NanoKVMs (Go and Go+). 2 physical @OlaresOS-based PCs (@Hi_MINISFORUM & @Acemagic_MiniPC). 1 Tailscale network. I accessed both Linux consoles remotely from outside: independently of the host OS. All good - Mouse HID: nope. Next: AI-assisted bare-metal recovery.
1
139
#JensenHuang speaks in Berlin tomorrow :: slavasolodkiy.substack.com/p… My ears are still in London:( I was supposed to be there. I won't make it — but hearing him twice on 14 October left me with four questions worth taking into the next keynote :: linkedin.com/pulse/listening… These are my recollections and interpretations, not a transcript. @nvidia people and fellow attendees: please correct or add to them!:)
1. YOUR AGENT NEEDS MORE THAN A PASSPORT. Who created it? Who is it acting for? Who can change its role or revoke its mandate? What evidence travels with it into somebody else's system? Identity is not authority. Knowing who someone is does not authorise them to move your money. That's the question behind my AuthorityClaw experiment: solodkiy.cv/authorityclaw.ht…
2. THE NEXT AI EXPERT MAY BE AN EXPERT IN SOMETHING ELSE. I heard a shift from tools for creators and developers towards agents adapted to particular industries. Less “How do I become an AI specialist?” More “What important problem in my industry remains unresolved?” The customer doesn't wake up wanting an agent. They wake up wanting something to work.
3. YOUR NEXT PC MAY NOT BE WITH YOU. OR ENTIRELY YOURS. First you went to the computer. Then you carried it. Next, perhaps you simply reach it — from whatever device is handy. That resonates with my @Olares_OS experiments. My next question is sharing: perhaps ownership matters less than authorised access, privacy and control. solodkiy.cv/olares-seamless.…
4. THE AI FACTORY NEEDS NEIGHBOURS, NOT JUST MEGAWATTS. At the evening event, I remember openness to different energy sources and infrastructure integrated with nearby communities. That connects to a hypothesis I'm developing: a Power Node combining electricity, compute, grid flexibility and useful local services. solodkiy.cv/power-to-intelli… My Hoxton Jobcentre concept asks what the last part could mean for people looking for work: solodkiy.cv/hoxton-jobcenter… What happens outside the fence matters too. Four subjects. One recurring question: WHO GETS TO USE THIS CAPABILITY, ON WHOSE AUTHORITY, AND FOR WHOSE BENEFIT?
Thanks to Elliot, Ayesha, Mozhgan and Manel for London and the Berlin invitation! Sorry to miss the latter:( My curiosity hasn't cancelled its trip. @NVIDIAAI's own London recap: blogs.nvidia.com/gtc-berlin-… Those listening tomorrow: what carries through, what changes, and what did I miss? My session shortlist is in the comment below. What question are YOU taking into the room? Keep asking linkedin.com/pulse/listening… slavasolodkiy.substack.com/p…
If a model can read an invoice, who is allowed to let the money move?
Yesterday my #AuthorityClaw solodkiy.cv/authorityclaw.ht… (built on Olares One) was selected first place in the @NVIDIA Claw Agent Challenge: DGX Spark Founders Edition, a GTC Berlin pass, and a live demo at Build-a-Claw on 14 October, London.
Not a research prize. A city developer challenge, remote, open to anyone in the UK. I do not know how many entries there were, so I will not invent a denominator. What I can say is what ran.
The chain on the bench:
→ #Nemotron as perception, not authority. It read synthetic passports. A deterministic MRZ gate and a #Prolog policy refused malformed extractions. NEED_MORE_EVIDENCE. The model explained. It did not override.
→ The problem, mapped. Banking onboarding, correspondent payments, digital identity. The question underneath all of them is the same: on whose authority is this agent acting?
→ #NemoClaw 1.0.30 was on the same @Olares_OS One (bundled v0.0.38, OpenShell 0.0.36). Useful as an orchestration shell. For the authority-critical path I did not trust the shell to be the gate. The verified run used AuthorityClaw’s own persistent Python harness, calling the local model directly, plus an OpenClaw-compatible authority-gate skill. That is a measured no, not a dismissal.
→ The pilot. Fictional company, simulated Faster Payments rail, no real money moved. Six invoices. A live revocation. A bank-detail change. Two prompt-injection attempts inside invoices, treated as data.
→ A quantum double-check on a separate job, reported with the mismatch it actually produced. Engine output is evidence, not interpretation.
The ladder only narrows:
→ Northside Roasters Ltd → Jane Okafor, director, £25,000 / £50,000 a week → Tom Hale, office manager, £1,500 / £5,000 → the payments agent, £1,000 / £3,000
→ Jane revoked Tom mid-run. The next ordinary invoice was refused. The inherited mandate no longer stood. £1,240 was held for a human. £4,900 to an unapproved supplier was refused, including the instructions the invoice tried to give the agent. New bank details on a known supplier stayed held for out-of-band confirmation.
= The model reads. The gate decides. The rail executes only against a valid receipt. AI override = false.
Box: Olares One, @NVIDIAAI RTX 5090 Laptop, 24 GB, unsloth/Qwen3.6-27B-GGUF Q4_K_M through llama.cpp. Seven receipts, hashes on the page, chain intact.
This does not make it a payments system, or KYC, or regulatory approval. First place validates the problem-to-product-to-engineering approach, not compliance. Real deployment still needs key management, a human review UI, audit export, and a lawyer on the mandate semantics.
I am not a career software engineer. I founded a bank and took a de novo licence through its regulator. The interesting part, for me, is that the same question can now be an executable artifact rather than a diagram.
Page Repo: github.com/slavasolodkiy/aut…
After completing the @Google × @NVIDIA learning pathway, I extended my #Sovereign Decision Plane #AI experiment with a real @GoogleCloud Storage witness: linkedin.com/posts/vsolodkiy…
A 545-byte receipt from a recorded synthetic decision was uploaded to a private bucket, then downloaded from the exact object generation. The bytes and SHA-256 matched.
NVIDIA #Nemotron provides candidate extraction in the recorded #localAI experiment; verification and policy determine the downstream decision. The local decision core is unchanged.
This demonstrates receipt storage and byte integrity—not certification of decision correctness = but the #cloud additional layer could give you much better flexibiliy, for example, for inter- and correspondent- banking communications (and GDPR\similar regulation to split local vs cloud data flows)!
#NVIDIAGTC @Jennifer_Dee
if two boxes can be clustered, which of the three jobs are you actually asking them to do? The Framework Desktop 495 (@FrameworkPuter is the same 192 GB cup as the @Peladn_MiniPC, btw, the same straw, the same #GorgonHalo silicon) — and a completely different sermon. Repairable platform, real clustering paths, honest documentation. Also a soldered heart. Candidate for pre-order: prosumer.watch/dossiers/fram…
□ Repairable is true for the skin. Swappable front tiles, expansion cards, standard fasteners, a pre-installed Noctua 120 mm. The exception is the expensive part: the @AMDRyzen AI Max+ Pro 495 and the 192 GB are soldered. Framework’s own spec page says so. The parts you would most want to swap are the ones you never can.
□ The open-ended PCIe x4 accepts a long card — an @Intel E810, 50 GbE, RDMA — and is not brought out of the standard case. Possible on the board is not convenient in the box. Plan the chassis before the card.
□ 120 W sustained, 140 W raised, is the CPU’s rating, not a wall measurement of a 4.5 L machine with fans, NVMe, a NIC and losses. Never copy a chip’s TDP into an electricity-bill row. Pre-orders opened September 30. Shipments claimed for November.
“Can I cluster two?” is three questions wearing one coat - Framework itself answers them differently:
→ Different models or agents in parallel:: The friendliest case. USB4 point-to-point went from about 9 Gb/s to about 30 Gb/s after an @AMD BIOS patch. Built for pipeline parallelism and multi-agent work. Most achievable, least glamorous.
→ One model that does not fit in one box:: Tensor-parallel territory: 50 GbE with RDMA, and real ROCm/RCCL labour. A published community build made a two-node Strix Halo cluster work with vLLM and an E810 — and describes the networking as serious work, not a checkbox. That testimony is the previous platform, not an independent test of the 495.
→ One model, answering faster:: Measured speedup on the specific sharding mode, not the sum of two spec sheets. “Linear speedup as you add nodes” is Framework’s sentence. From “the network is slower than local memory” you cannot conclude either scaling or its impossibility. The method decides what crosses the wire. The wire decides the bill.
Claims in pencil. Measurements in ink. Distributed inference in 2026 still bills in driver versions, NIC firmware and RoCE verbs. If your ceiling is one pasteable block, a two-node tensor-parallel lab will find it and raise it a career.
Same cup-and-straw as the Peladn underneath: about 6.7% more theoretical bandwidth, 64 GB more capacity, residents rather than speed. Always-on companion is a no — this is a 4.5 L desktop, not the 30 W pet. (Eight tests stay mostly unsigned until a box ships.) The Noctua is encouraging. Encouraging is not a dB(A). Not fiction, and not plug-and-forget. The cluster is real. The easy is workload-dependent. The evenings belong to ROCm.
Surface? Cute toy. #GPDPocket4? Pocket-sized BEAST crushing it in power for the SAME cash! Ryzen smokes Snapdragon – games, AI, raw muscle. Who needs it when you got portable firepower? 💥#TechRebels Get -50$ discount with 'DRAMGOLD' on @GPD_Official here: nextdoorai.medium.com/gpd-as…
133
when a headline says 192 GB RAM, 160 GB VRAM and 24 TB, which of those three is actually in the box? The Peladn YO2 495 @Peladn_MiniPC is a candidate filing — what the silicon is, what the headline inflates: prosumer.watch/dossiers/pela…
□ 192 GB and 160 GB are one pool of LPDDR5X, not two products. Up to 160 GB can be allocated to the @AMDRadeon 8065S. Allocation is an invitation, not a velocity. The straw decides how fast the invited models drink.
□ 24 TB is the wardrobe, not the clothes. Three M.2 2280 slots, about 24 TB at the ceiling (and the described offer ships a 2 TB SSD. ¥49,999 on JD, 192 GB plus that 2 TB.) China-first, pricey, and still a real point on the map!:)
□ The 495 is #GorgonHalo: 16 Zen 5 cores, 40 RDNA 3.5 CUs, up to 160 W. Same CPU and GPU architecture as the @FrameworkPuter Desktop 192 GB tier.
The JD card behind the write-ups did not render a readable spec table. So the parameters stay candidate, as reported by Notebookcheck and the platform notes, not as a verified receipt. Promotion needs the receipt. That is the pencil rule, not a snub.
Why file it anyway: (a regional, unfavourable box is easy to dismiss, and dismissing it is how) this AI-first mini-PC quietly becomes a list of favourite devices:) The Peladn marks where the 192 GB tier is going, and at what price. It does not change the picture of the platform.
Capacity is not speed. Ok, a 24 GB GPU cannot keep a giant MoE home beside the desktop, the files and a VM. 160 GB allocatable means the biggest open models simply live on the machine — slow to sip, always there. Do not buy it for speed. If the question is how fast it answers, a mid-size dedicated GPU wins at half the price. Confuse the two and you return a perfectly good machine (for the wrong reason)!
The honest lever is storage. Three slots, the ceiling real, the clothes not installed. The RAM is soldered. Buy the cup you will need in year two, at year-one prices.
Eight tests stay mostly unmarked. Kettle stays cold until a reader reports. Always-on companion is a no — this is not the 30 W pet. Minute twenty on a library desk is unsigned. Filed as a wish, wanna-be-loved.
The @nvidia Innovator's Dilemma' :: dram.gold (New Book!)
ISBN: 9798235142671 (e-Book) ASIN: B0GZ42SBNY
ISBN: 9798195009434 (Paperback)
DOI 10.6084/m9.figshare.32133316
on @amazon & @eBay #InnovatorsDilemma #AI #NVIDIA #JensenHuang @NVIDIAGeForceUK @NVIDIAGeForce
71
if a box has two memory pools, do you add the gigabytes — or do you ask which part of the job each pool is allowed to do? The Lucebox Zero 495, @luceboxai by @pupposandro, is the first machine in this series where the interesting part is not the CPU generation. Two pools, one placement engine, a PCIe pipe between them prosumer.watch/dossiers/luce… Wanna-candidate! Ships January 2027. Not a review:
□ Do not sum. @AMDRyzen AI Max+ Pro 495 with 128 or 192 GB LPDDR5X at 273 GB/s, and a @AMDRadeon AI PRO R9700 with its own 32 GB GDDR6 at 640 GB/s. 192 + 32 is 224 GB of memory. It is not 224 GB of fast VRAM. Two sports, one chassis.
□ The Engine is the actual claim. Hot experts and the always-used parts go on the R9700. The rest live in Ryzen memory, where they are executed, not parked. Grouped work, batched transfers, the merge moved from CPU to GPU. That is software placing compute with the data. It is also, so far, the manufacturer’s own description and numbers.
□ “Faster than DGX Spark” was run on the previous 395 build — different settings, compression and test modes. Useful signal (wrong generation?). Zero 495 gets a clean sheet until its own signed runs.
The buyer asks how many GB per pound? Ok, the research question is what other compute system can be built from this shape:
→ Weights and experts:: Interpretation, candidates, the LLM’s day job. Heterogeneity is an advantage only if you stop forcing both hearts to do the same work.
→ Facts, roles, relations:: An authority check that does not re-read the whole history as text. Keep ALLOW next to the facts, the rule version and the mandate state it came from. When the mandate dies, revise the touched conclusions. Reused memory must never become reused stale permission.
→ Search states and deltas:: One fact base, many changes: mandate expired, director changed, delegation capped, evidence voided. The useful output is not “we ran more agents.” It is the minimal change that makes this action invalid, with a checkable explanation.
“GPU proposes, CPU proves” is a habit, not a law. Which parts of a delegation chain are really bulk relation operations is an experiment (and a purchase argument?).
Outside, the sermon is the Olares One prosumer.one/guides/olares-o…: Ubuntu Server, setup from a phone, management in a browser, apps pointed at a local API, signed updates with a health check and automatic rollback. Complex infrastructure inside, one short path outside. A pretty page does not prove the box never becomes a ROCm admin job. That is the Kettle test, and it is the whole product question.
The rest of the receipt, filed as an offer:: 340 × 110 × 320 mm, 6 kg — almost twelve litres with handles, a compact workstation, not a cube. About 500 W under load, 40 W idle, a 1,000 W supply that is a ceiling, not a diet. None of those watts are signed. UK prices with VAT and duty printed in pounds: £5,999 for 128 GB, £7,999 for 192 GB, first 200, base later quoted at £6,999. The saffron line is +£2,000 for 64 GB. A basket saved on 5 October — 192 GB, 4 TB, Mellanox CX556M at 2×100 GbE — came to £9,048. Fair.
Heterogeneous memory is not an automatic flaw. Unified memory is not an automatic virtue. The question is what useful possibility the architecture opens (and at what price).
Nice one, @TiinyAILab × #Odysseus = demo is one of the more practical things I’ve seen lately in local AI!
I’ve been running a mix of @Olares_OS + @gmktecdotcom + @softwincn's GPD Pocket 4 + eGPUs and various mini-PCs for a while. Ordered two Tiiny units specifically for this!
135
if the GPU is in another box, which bytes actually have to come home? WiCi One by @zili_meng is not an @AorusOfficial with the cable replaced by Wi-Fi: prosumer.watch/dossiers/wici… The AORUS question is simple: my mini-PC gets a powerful GPU. The @WiCiHQ question is different: my devices get access to a shared GPU without stopping being the devices I already use. Candidate-to-buy 100%!
The slogan hides the load-bearing wall. What stays by the GPU, what crosses the wire, and how the software hides the split from the app. The radio is just the radio. The software layer is the product.
□ Physically it is a standalone node — its own CPU, storage and OS — not a graphics card with a Wi-Fi module bolted on. Novelty lives in how capabilities are presented to applications.
□ One brochure promise, three documented modes. API: the app asks a local service and gets a result. Useful, and not a wireless graphics card. SDK: the developer aims work at their environment, wici.gpu. Driver: an existing GPU app is supposed to see a virtual accelerator and run without a rewrite. A working chat proves API.
□ Their own technical article is more honest than the shop. Active data in VRAM, saved state in the box’s RAM, model weights in an SSD cache next to the GPU. If the next operation needs the last tensor, do not send it home just to send it back. The client can hold a logical id. The interesting question is not whether Wi-Fi can carry PCIe’s bytes. It is what fraction of the exchange can be skipped entirely. Wi-Fi is not PCIe — their words. Some tasks are too latency-, data- or driver-sensitive. Respect!:)
One box, many devices is declared. The family route already running here is a different way, not a worse one: the game lives on the machine, the other person reaches it through a browser. Driver mode, if it matures, would leave the app on the laptop and send the GPU-work away. Two routes to the same human wish. Neither is automatically the winner.
At $9,999 for the @nvidia RTX 5090 32 GB edition — $2,599 for the 5060 Ti 16 GB — you are not buying a GPU. You are buying the integration tax back. Ships in 8 weeks is a window (not my witnessed delivery, yet). A $600 pioneer discount does not change the shelf, but cool. NYC shows on 5–6 October are a chance to see the claim outside one promo film, btw!
Still not an independent result\tests. (Examples on the technical pages are marked preliminary.) The checks that would promote it are narrow. A named app, a named OS, a named mode. What changes after the first launch, through a wall, with a second user attached. Whether results and data separation survive memory reuse and a dropped link. For the compliance line, the research question is ours, not their feature: can user context and authority stay on the client, while compute is delegated with the minimum necessary data?
Backpack is a no = Meeting room is the fascinating case: the GPU sits in the office while the laptop visits the client, and data separation across that link is the unsolved check. Idle watts of a 5090-class shared appliance are unpublished, it's ok. Best of luck in NYC!
My #PerfectInfinityCraft card dram.gold/aorus.html redesign note from my side!) #GIGABYTE40INFINITY @AorusOfficial's member: @slavasolodkiy
148
#GhostCore @ghostai is not a new record in memory architecture. It is a bet on the finished personal AI product — a helper that remembers your life across model changes, acts on your behalf, and never asks you to become its administrator. Filed as my next wish-candidate: prosumer.watch/dossiers/ghos… (not a review!) Ships October 31. The landing page says it beautifully. The privacy policy says it more precisely. Read both:
□ The metal is modest on purpose. RTX PRO 4000 Blackwell SFF, 24 GB GDDR7 ECC at 432 GB/s, a 70 W card — @nvidia's number for the card, not the box. @AMDRyzen 5 7600, 64 GB DDR5, 1 TB. $3,499. Nothing here is trying to win a benchmark. A compact always-on node, not a peak-power monument.
□ $11M, led by @a16z, says the category has investors’ attention. Capital validates the market question.
□ Preinstalled weights are everyday models, not a 120B tin. @HuggingFace and your own weights are the claimed door. 24 GB is a cup. Giant MoE residents are a no.
The loop is the product, and the arrows are the exam.
→ Sources feed memory. Apps, files, devices, and — in the policy — browser passwords and live sessions. Enormous power, enormous blast radius.
→ Memory is supposed to outlive the model. Change the model in settings; context kept, nothing re-uploaded. Saving records to disk is easy. Transferring meaning, provenance and access scope between minds is the hard part. A stored assumption must not quietly become a fact.
→ Actions are where data leaves. Conversation and memory run locally. Online actions can send what they need via Ghost’s gateway or directly. Ghost not storing it does not mean the end provider does not. Own API keys are supported. Remote access can be replaced with your own link — Tailscale named explicitly. Those are the sentences of a company that has read the same problem.
The revocation test is the one I actually want!:)
An agent reads a document under a work mandate, extracts facts, then the mandate is withdrawn. Disconnecting a source does not delete saved data. Deleting the file does not necessarily delete the memory made from it. What happens to the original, the derived memory, and the permission to keep using the conclusions?
“The agent can access my account” is not “I authorised any action through it.” Read the invoice: fine. Forward it to a third party: not fine. Both use the same legitimate service. The firewall @zainmfj described — monitor every outgoing request, reject a sanctioned action or an exfiltration, flag it — is the deciding feature, and it is still in development. I would not score address-blocking. I would score a concrete action bound to a concrete permission, with expiry and revocation.
Eight tests stay mostly unsigned. Kettle looks like the shortest route from box to helper in the series, if the guided setup works as written. Reading is the honest middle: comfortable everyday models, no fireworks. Backpack is a no — this is a home appliance, not luggage. Edinburgh train splits: local chat, actions that leave. Meeting room is the dream, and the proof is what exactly leaves the room when the agent acts.
The high score waits (not for another demo chat, but) for proof that a normal human gets useful work, understands the memory, restricts the actions — and survives the mistakes. A working note. The slogan says never. The policy says when. My next wish though!:) and - congrats for the team! looks awesome
waiting for my @TiinyAILab prosumer.one/guides/tiiny-po… delivery - and counting days!) as a kid on Christmas Eve, hahah
147
Three Ways I Tried to Turn a Mini-PC into a Simple Remote Gaming Machine
Over the past couple of weeks I tried three different ways to make a small PC behave like a simple “open The @TheSims 4 and play” box (for my wife).
The surprising result: the technically most sophisticated solution was not the easiest one to live with.
1. @SteamOS / @bazzite_gg on the @Acemagic_MiniPC AM18 Cube
The first idea was the obvious console-like path: install a gaming-focused #Linux distribution and make the mini-PC behave more like a Steam Deck or living-room console.
I experimented with #SteamOS/#Bazzite on an @Acemagic_ AM18 Cube.
This is attractive in theory:
→boot directly into a gaming environment;
→@Steam-native UX;
→controller-friendly;
→no need to expose a general-purpose desktop.
But in practice it still felt like maintaining a dedicated gaming machine. Hardware support, GPU choices, @Linux quirks and the fact that this was only one of several mini-PCs in the house made it less compelling than I expected.
It worked as an experiment, but it did not solve the bigger UX question:
How does a non-technical family member simply open a game from whatever device happens to be nearby?
2. Sunshine + Moonlight on #OlaresOS
The second path was much more interesting technically.
I installed Steam Headless on OlaresOS, first experimenting around the Acemagic setup and then moving the working setup to the @iMoreFine M900.
The stack became:
OlaresOS+
→ Steam Headless
→ Sunshine
→ Moonlight
→ iPad / another client
Eventually it worked very well.
But getting there involved more moving parts than I expected:
/separate SUNSHINE_USER / SUNSHINE_PASS;
/Sunshine pairing;
/Moonlight PIN pairing;
/Olares Overlay Gateway;
/manual LAN host setup;
/direct ports for video, audio and control;
/debugging a “No video received from host” error;
distinguishing multiple auto-discovered steamheadless-* instances.
The key fix was Olares Overlay Gateway, which gave Steam Headless a direct LAN IP and allowed Moonlight to connect to the correct local endpoint.
Once configured, the performance was good.
But for our real family use case there was another practical limitation: in our setup, Moonlight was happiest when the client was on the same local network and VPN was disabled.
That is fine for a gaming enthusiast. It is less fine for someone who just wants to play The Sims 4.
3. The simplest solution turned out to be the @Olares_OS web-interface-management = the winner was the least exotic path:
On the same Morefine M900 running OlaresOS, we simply used the browser-accessible Steam Headless desktop.
My wife logged into the Olares desktop once from her Chromebook.
After that, the workflow became essentially:
open browser+
→ open Olares Steam Headless
→ play
No Moonlight pairing.
No need to remember Sunshine credentials.
No LAN IP.
No dependency on being on the same Wi-Fi.
No requirement to disable VPN.
And, importantly, it works from different devices and networks.
For her actual use case, this was dramatically better.
What I learned:
→The technically “best” architecture and the best user experience are not always the same thing.
→SteamOS / Bazzite is appealing if you want a dedicated gaming appliance.
→Sunshine + Moonlight is excellent if you care about low latency, direct LAN streaming and gaming performance.
=But Olares’ browser-accessible desktop turned out to be the most practical option for a normal household user who simply wants to open a game from a Chromebook, laptop or phone without understanding the infrastructure underneath.
That distinction matters.
For a power user, the stack is:
GPU+
→ Steam Headless
→ Sunshine
→ Overlay Gateway
→ Moonlight
For my wife, the ideal stack is:
Browser →
→ Olares
→ Steam
And that is probably the more important product lesson.
The infrastructure can stay complicated.
The user journey should not be.
If I were designing the ideal Olares gaming experience from here, I would keep Moonlight/Sunshine as the high-performance option, but make the default path feel much closer to:
Open Olares → Play Game
That is the version people will actually use. // sic @TheSteamOS @Steam0S @SteamOSLinux
Tried to DIY a #PistonV analog: AceMagic AM18 Cube #1 on Bazzite, #2 on SteamOS. You can assemble the stack. The result is not the same.
Same with the circular screen. I tried using my Kickstarter-backed Decokee as a stand-in. Technically possible. Not easy. And at the actual user level, it is not the same product. That integrated 4" boot/status display is the point.
Closest Kickstarter cousin was TECNO TAURUS / MEGA MINI G1 Pro — now retail, and it just won the IFA 2026 Gold Award for Miniaturized Gaming PC Technology. In some ways Piston V is already more interesting: gadget form factor, dual-OS tap-to-boot, living-room console feel. On eGPU it could have competed harder if Kickstarter actually bundled concurrent campaigns — say with Orcanyx (funded; hoping they make it through production and delivery) or eGryphon (funded; then production/delivery appears to have collapsed).
The value of Piston V is not “a mini PC + a Linux ISO + a round screen.” It is the finished object. kickstarter.com/projects/pis… on @Kickstarter
1
358











