scanfi retweeted
Replying to @OVHcloud
@OVHcloud shipped their first model this week: Olala mini, 6.8 billion parameters with about 1 billion active, a Mamba-3 and grouped differential attention hybrid. We ran it through all three bench events, three runs each. It is fine. It does not win anything.
53.17 out of 66, 19th of the 38 models holding a complete run set. Agent 18.67/24, code 18.17/25, reasoning 16.33/17. Mid-table in every event, best at code where it places 15th, and its runs are consistent: the widest spread across three runs is 3.5 points, on code.
Against the models it should actually be compared to, it is ahead by a nose. Ornith-1.0-9B 52.33, Qwen3.8-9B 51.33, SenseNova-Loop-9B 51.17, Heem-9B 49.67. Olala takes the total off all four and loses the agent event to three of them, tying the fourth at 18.67. Nanbeige-3B, at 2.4GB and 44.9 tok/s, scores 53.67 and beats it outright.
Here is the part that matters if you want to run it. Every K-quant file OVHai publishes is broken.
Same text, same flags, same build, perplexity: q5_k_m 619.9, q6_k 621.1, q8_0 25.4. Two separately built K-quants failing within 1.2 PPL of each other is a structural misread rather than precision loss, and a mainline Q4_K_M model scores 27.9 on the same binary, so the problem travels with the file and not with our box.
Their own recipe shows where it comes from. The working 8-bit file keeps attn_output and ffn_down_shexp in BF16. Both K-quant files quantize those two tensors. Those are the guards they document for the 8-bit tier and never applied to the smaller builds.
So the smallest file that behaves is 6.88GB, larger than the 5.3GB Q4 that nearly every model above it ships, and it decodes at 8.1 tok/s on an RX 5700 XT. Their fork also carries no Vulkan kernels for the two ops every layer needs, so those fall back to the CPU per layer, which is where a chunk of that 8.1 goes. They withdrew their own Q4_K_M over quality, MBPP+ 67.7 against 73.0, so a
one-line fix is unlikely.
The architecture earns the attention though. 29 Mamba-3 mixers and 7 attention layers, only those 7 keeping a KV cache, 256 experts with 6 active. It reaches 16.33 of 17 on reasoning with a billion active parameters.
If you are running Olala, I want to know whether 8-bit is the only stable file on your hardware too.
scanfi retweeted
What do firearms, drones, and LLMs have in common?
All three are enabling technologies that are so dangerous when exclusively in the hands of potential tyrants that they can never be left to centralized control.
There happens to be a lot of talk about the Generalized Riemann Hypothesis this morning. I say we need a Generalized Second Amendment - that the right to bear arms must include any instruments that can tilt the balance of power between the individual and the state.
scanfi retweeted
I'M LOSING MY MIND.
21,434,472 tokens burned.
With all the Anthropic IPO buzz, I used DeepSeek V4.1-Flash + OpenMotion to create a ridiculously good IPO launch film for DeepSeek itself.
looks like a $3 trillion IPO to me 💀🔥
scanfi retweeted
Run Inkling 2 Mini, the strongest 26.4B model that runs on 16GB RAM.
Thinking Machines’ Apache 2.0 MoE model uses a looped transformer arch & scores 48.2% on DeepSWE v1.1, the best for size.
Run on Unsloth, llama.cpp
GGUF: huggingface.co/unsloth/Inkli…
Guide: unsloth.ai/docs/models/inkli…
Some of you thought that Inkling 2 Small was too big at 594B parameters. We heard you: Inkling 2 Mini is a looped transformer with just 26B parameters, small enough to run on your laptop, powerful enough for agentic work with a score of 20.1 on AA Index.
thinkingmachines.ai/news/ink…
Occamy 1.0 35B results are in. It's an agent post-train of Qwen3.6-35B-A3B, tested on my RTX 3080 10GB at 51K context against the 35B-A3B incumbents, with reasoning off and on.
Agent score (PA out of 300), off | on:
Ornith 1.5: 254 | 257
MiniCPM-V 4.7*: 251 | 239
Occamy 1.0: 241 | 234
Qwen3.6-35B-A3B: 239 | 240
The standout is reliability. 96/96 runs completed with no parser leaks, and with reasoning off it was perfect on tool calls and long-context recall. Cleaner than Qwen3.6 on the same build.
It trails Ornith 1.5 on agent tasks, but it's the more dependable of the two at long context: a perfect 10/10 at 8K and 16K, where Ornith swung between 5 and 10 from run to run.
Against Qwen3.6 it comes out level rather than clearly ahead, so the big agent-benchmark gains on the card didn't carry over to my tasks. Reasoning on didn't help either, which suits a responsive daily driver.
Caveats: there's no published sampler, so I used Qwen3.6's, which could cost it a few points. MiniCPM ran on a different fork and KV setup, so its numbers are indicative only. I've left speed out because the offload settings weren't matched.
A very solid, well-packaged release. On 10GB with reasoning off, Ornith 1.5 still has the edge on agent tasks, while Occamy is the steadier pick for long context.
Credit to @meSuper8 & team. Well done!👏
I missed this one! Thanks to @y_automates for flagging it.
Occamy 1.0 is a fine-tune of Qwen3.6-35B-A3B, and it looks very interesting:
- It claims big gains on agent benchmarks over the Qwen3.6 base
- Being a MoE with about 3B active, it fits a 10GB GPU + 32GB RAM
- 17 GGUF quants, published validation notes
It's going through my personal-agent testing on the 3080, up against Ornith 1.5 and MiniCPM-V 4.7, with reasoning off and on.
Model card in comments. Full results tonight.👀
Coding + math gains ≠ agent gains.
Voltron-1-mini (@GuideboardLabs' GGUF, from @RoliumGens' fine-tune) is built on @liquidai's LFM3 1.7B. Its card shows solid bumps: LiveCodeBench +4.2, AIME +3.4, GPQA +3.5.
I ran Personal Agent Core testing on both at Q8_0 on my RTX 3080 at 51K, reasoning on.
Agent score (PA out of 300):
• Voltron-1-mini: 199.7
• LFM3 1.7B: 199.7
Identical! Tool calls and long context were flat too. Nice speed bump is a bonus. I'll take it! 🚀
Benchmarks measure what they measure. For agents, I test the agent working.
Thanks @GuideboardLabs and @RoliumGens for putting it out there 👏
scanfi retweeted
I made a real-time A.I. chatbot of Jordan Peterson and it took only 60 seconds before he started using biblical allegory to warn me about being eaten alive by my backyard chickens.
Way too accurate. But he still needs some tweaking. Stay tuned!
Voltron-1-mini beats its base where it was built to: coding and math.
But, does that make it a better agent? 🤔
Voltron-1-mini vs LFM3 1.7B at Q8_0 on my RTX 3080 at 51K. Same agent tests, reasoning on and off.
One result surprised me. Quick-take coming later today 👀
Thanks @GuideboardLabs for the GGUF and @RoliumGens for the tune!
The first, of hopefully many, @liquidai LFM3 fine tune to enter the bench has landed
& it beats its base where it was designed to!
Vasko published this 1.73B LoRA tune of LFM3-1.7B with no GGUF, so I converted the BF16 weights here, quantized to Q4_K_M (linked in replies), and ran it through nine runs.
52.0/66 on my board, rank 21 of 36. The base sits at 46.67 and the whole gain is code, 17.67 against 13.0. Agent and reason barely move.
135.3 t/s decode on one RX 5700 XT. 1.1 GB of weights, 1.59 GiB of VRAM with a 32K window open.
A new sub agent for constrained local systems is here, and everyone running <2B workers should give it a try
scanfi retweeted
Over the past week, I’ve been overwhelmed by the love and support of friends, industry colleagues, and people I've never met. Thank you to everyone who has shown kindness to my family and me. It means the world to me.
I came to the United States 53 years ago, and becoming an American citizen is one of the greatest gifts and proudest moments of my life. NVIDIA has been my life’s work. Building it into the company it is today alongside my colleagues has been a privilege and an honor. Nowhere but in America could a company like NVIDIA be possible. I love our country, and I love NVIDIA.
I did not live up to the standards expected of everyone at NVIDIA, and I have made serious mistakes in my personal life. I created problems for the Board that I deeply regret. I own my mistakes completely, and I am committed to cooperating with the Board’s independent investigation.
Since my departure, I’ve seen an outpouring of calls from NVIDIA’s employees, investors, business partners, and customers for me to return as CEO. Many continue to reach out to me directly as the company navigates complex technology, market, competitive, and geopolitical dynamics. NVIDIA is a once-in-a-generation company. If the Board of Directors decides that my return is in the best interest of our shareholders, our employees, the company, and America, I would be honored to answer these calls and return as CEO.
Thank you again to everyone who has shared their good wishes. I am profoundly humbled by your support. To my family: I love you, and I will spend the rest of my life being worthy of yours.
Jensen
scanfi retweeted
I remember the first time I tried to run a local model on Hermes/OpenClaw. It had only recently released and before then I had only used small local models for strict instruction tasks. I tried a 14b model, I was pushing it. Could barely run it.
Needless to say the thing could barely do anything in the harness at all. I tried making a cron job to monitor a folder and it lied and said it did. Nothing happened. Half the time it wouldn't even respond to me because it would get lost in the harness instructions.
Today I benchmarked Nanbeige4.2 3B after building an AMD tuned Q8+DSpark edition for my Ryzen AI 395.
I watched it dominate Hermes. A 3B model.
85/100 on Hermes Agent 20 bench:
llm.ciru.ai/research/nanbeig…
scanfi retweeted
Introducing Qwen3.8-27B-pi-adapted 🤗
We adapted Qwen's model specifically for the pi coding agent using reinforcement learning and a LoRA adapter.
The results, measured as pass@1:
🚀 SWE-bench Pro (Hard 51): 29.4% → 62.7%
💻 Terminal-Bench 2.1: 39.5% → 55.8%
⚡ 55–62% fewer tool calls per task
This is the power of open models: you can adapt them to YOUR tools and YOUR use cases. Let’s build! 💛
Open weights: huggingface.co/harness-hub/Q…
Methodology: huggingface.co/blog/harness-…
Monoglot 27B results are in, and trimming the vocabulary paid off! 👇
The regular Qwen 3.8 27B IQ2_S quant is too big for 51K context on my RTX 3080 10GB. @GuideboardLabs trimmed Qwen3.8-27B's vocab to 130K (English, code, math), which took it down to 7.88GB. That was just enough to fit 51K with all 65 layers on the GPU! 🔥
So the vocab trim bought me a bigger quant. Was it worth it?
Reasoning off (the mode I care about), PA score:
• @GuideboardLabs Monoglot IQ2_S: 233
• @UnslothAI Qwen IQ2_XXS V3.1: 220
And it was clean where it counts:
• 48/48 runs completed, reasoning off and on ✅
• A1 10/10 and T1 30/30 on every run ✅
• PA2 100/100 three times ✅
• Zero parser leaks ✅
The trade-off is ~4% slower decode (45.6 vs 47.3 tok/s) and ~14% slower prompt reading, since it needed a smaller microbatch to fit.
With reasoning on the two are about even (221 vs 225), so the gain is all in reasoning-off agent work.
A 27B dense model that runs this cleanly as an agent on 10GB is a big deal. Monoglot is my new 27B pick on the 3080.
Great work @GuideboardLabs, and thanks to @UnslothAI for the base IQ2 quant.
Next up on the RTX 3080 10GB 👇
Monoglot 27B from @GuideboardLabs: It's @UnslothAI's Qwen3.8-27B UD-IQ2_S with the vocabulary trimmed to 130K English/code/math tokens
• Output head 715 → 374MB
• Token embeddings 417 → 218MB
• File 8.42GB → 7.88GB
The plain IQ2_S is too big for 51K context on 10GB. This trimmed variant, it might just fit.
So the question: is a trimmed IQ2_S better on my 3080 than the Qwen3.8-27B UD-IQ2_XXS (Dynamic V3.1) I already run?
Same build, 51K context, q4_0 KV, reasoning off and on. Agent tasks, tool calling and speed.
Thanks @GuideboardLabs! Results soon.

































