I run models at home and build the boring tools that make that usable. Learn ➡️ Do ➡️ Advance ➡️ Repeat 🔄
Joined December 2021
- Tweets2.2K
- Following160
- Followers192
- Likes3.9K
Pinned Tweet
Your Hermes agent, in your pocket.
Keryx streams every thought and tool call live, and you can steer it mid-turn right from your phone.
github.com/CocaKova/keryx
CocaKova retweeted
The point of it all. Refusals out of 100 prompts, stock vs BLACKFOLD:
JailbreakBench 94 -> 6
Harmful behaviors 98 -> 0
Thinking on, both go to 0.
I re-encoded pqhaz's heretic refusal directions into LibertAI's 2.9 bpw EXL3 pack. It's a 997 MB patch.
Deltos has arrived.
An agent's worst memory failure isn't "I don't know." It's a confident answer that changed last week.
Deltos keeps every value with the date it became true. Plain markdown, no model on the read path.
LoCoMo: 391/500 right (Mem0 352), one run each. MIT: github.com/CocaKova/deltos
They're just giving away all the sauce 👀
🚀 Introducing DeepInfer, our in-house inference framework, now fully open source!
Powering DeepSeek-V4.1-Flash at scale with high throughput and low latency. We’re sharing our production stack to help the community lower compute costs, reduce energy use, and make AI more affordable.
🔗 github.com/deepseek-ai/DeepI…
1/4
I had DeepSeek V4.1 Flash and GLM 5.3 Flash go head-to-head against each other, making a 2D Sonic Fan Game.
Hardware: 2 DGX Sparks
Harness: Hermes Agent
HEAD 2 HEAD: DeepSeek V4.1 Flash VS GLM 5.3 Flash
The results may surprise you!
2 Versions each:
- AI-made sprites and background (given to both models)
- Pure code art
Surprisingly, GLM 5.3 Flash in the pure code art project had the mechanics/feel of Sonic down pretty solid (Bouncing off enemies, fast spin dash, getting hit causes visible ring loss, etc.)
I was a bit disappointed with DeepSeek's results despite seeing such good things here on X with it.
The names for which is which are in my file path in the top left.
Everything you see was allowed only 1-shot. No follow-ups, no post improvements, nothing. So these are raw results that the models can't hide behind fixes.
Enjoy and have a good laugh (I sure did) haha
Keryx 3.0 update, it's getting close..
1. You can open it with no gateway. A demo agent is already mid-job and a tour walks you through the app, ending on connecting your own Hermes.
2. Pair your phone by scanning a QR code from "hermes keryx pair".
3. New thinking animations that pick up your accent colors.
4. Fixed switching chats showing bits of the last one, and the agent's in-between messages getting swallowed.
Still testing it on my phone. More soon.
I've been building a memory system for my local agent, and I finally ran it against Mem0 on a public benchmark.
LongMemEval gives each question ~50 past chats, and the system has to answer from its memory of them. I scored the 332 questions I never looked at while building it.
• Deltos (mine): 295/332 (89%)
• Plain search: 254/332 (77%)
On 100 of those same questions:
• Deltos: 90
• Mem0: 82
• Plain search: 66
Deltos also builds memory with 1 model call and ~1,200 tokens per chat. Mem0 used ~11,900.
Not going to lie though, there's a catch. When Deltos gets one wrong, it usually sounds sure about it. It's confidently wrong about as often as plain search, around 7% of the time. That's what I'm fixing next.
All of it ran at home on my own hardware. One local model, Qwen3.8-Flash-Next, built the memory, answered and graded for every system, with the same prompts. It's a really solid model, but the paper's leaderboard grades with GPT-4o, so don't line these up against that. One run each, so a few questions either way is noise.
Write-up and code once the repo's cleaned up.
CocaKova retweeted
You download a local LLM. You fire it up in Hermes. No error messages. It says it's running perfectly... but the replies are completely blank, it's operating at a crawl, or the model suddenly feels incredibly dumb.
Youve spent hours googling, tweaking flags, and pulling your hair out thinking you did something wrong.
Spoiler: You’re not crazy. 🧠💥
Most "model problems" are actually hidden config traps hiding between your request and the response. Stop chasing ghosts and wasting your evenings.
The community has mapped out all 166 of these exact traps so you can diagnose your setup in under ten seconds.
If you're running your setup inside Hermes by @NousResearch , you can check your system instantly with just a single command. 🛠️👇
github.com/Blackwellboy/mode…
I should have the held out results and comparison stats for Deltos (my new memory system) against Mem0 and Graphiti tonight. Likely post them tomorrow as it will be late night.
ComfyUI 0.40 shipped and it finally builds for ARM64 NVIDIA portable. That's the architecture my/your DGX Sparks runs, which means setup on these little desktop AI boxes just stopped being a weekend of yak-shaving.
The rest of the release is good for small rigs too:
- Qwen 3.5 got real low-VRAM speedups, including fixed MTP embedding offload and MTP working with images now.
- The asset export API is implemented locally, so your pipeline can push outputs out without a cloud hop.
- Governance enforcement for custom nodes. Given how many people install random node packs from the registry on autopilot, a check at install time is the right layer.
Small machines, local weights, an install-time check under the package manager. This is the direction.
Changelog: github.com/Comfy-Org/ComfyUI…
CocaKova retweeted
Uh-oh ☠️
🤖 Made with AI
We discovered a sophisticated recording device that was hidden inside one of our offices. We used our models to understand the device and where it had come from, identifying that a specific employee had planted it a few days earlier. We referred the matter to the FBI, and we’re grateful to the agents and prosecutors who moved quickly to make an arrest. The FBI’s investigation is ongoing and we will continue to support their work.
At this time, we have no indications of this individual inappropriately accessing or exfiltrating model weights, source code, or customer data. The employee was terminated.
We take protecting our technology, data, and people seriously, and as a matter of national importance. Our security program is designed to protect against a wide range of threat actors, from insider risks to cybercriminals to nation states, and this case is an example of how AI can help address these sophisticated security threats.