Mateusz Mirkowski (@llmdevguy) header
Mateusz Mirkowski (@llmdevguy) profile picture

Autonomous agents, agentic engineering Building & testing agentic systems Exploring local and Chinese LLMs

Remote work evangelist
Joined March 2013
Photo by Mateusz Mirkowski (@llmdevguy)Photo by Mateusz Mirkowski (@llmdevguy)Photo by Mateusz Mirkowski (@llmdevguy)Photo by Mateusz Mirkowski (@llmdevguy)Photo by Mateusz Mirkowski (@llmdevguy)Photo by Mateusz Mirkowski (@llmdevguy)Photo by Mateusz Mirkowski (@llmdevguy)Photo by Mateusz Mirkowski (@llmdevguy)Photo by Mateusz Mirkowski (@llmdevguy)Photo by Mateusz Mirkowski (@llmdevguy)
GPT-6.2 Astra High Nice or not?
1
8
233
It looks like for coding it's more reasonable to stick with Sol 6.2 instead of Astra.
85
How this can be so cheap? 😳
46
Another reason to switch to Omarchy.
macOS 27.3 Warns Against Giving AI Apps Full Mac Control macrumors.com/2026/11/16/mac…
48
What’s going on with OpenBot? Why it’s better than Hermes?
49
GPT-6.2 series are coming soon. 👀 Let’s crush Opus 5.5!
1
148
Kimi K3.1 Pro Abliterated: a coding model with censorship removed and no IQ drop. They cut just the refusal vector from the weights, without retraining or dumbing it down. Running it takes 8x B200 GPUs (1.5+ TB VRAM😳), while the API starts at $2.5/1M input and $5/1M output tokens.
⚡️ Introducing Kimi K3.1 Pro Abliterated, the most powerful open model for cyber and coding. • Native MXFP4 • 1M context • Zero data retention $2.5 input / $0.25 cache / $5 output
3
369
This is great model for simple automation tasks, but don’t use it as Hermes brain. Stick to Deepseek 4.1 flash which score in terminal 4.0 42.3% vs 19.2% for Hy4 Flash. Anyway “free” is reasonable price. 💪
Hy4 flash, the new model from @TencentHunyuan, is free in Nous Portal for two weeks. It is built to be extremely cost-effective at everyday work tasks like documents, slides, and computer use. Try it today at portal.nousresearch.com
2
183
Terra is still great, cheap code reviewer. Extra high often catch more bugs than Sol high.
Nobody uses Terra anymore idk why, it's a great sub-agent
1
111
This time Gemini Argon High. I really like the result! Consumed 10% of weekly Argon's limit. This makes mi feel something went wrong with Hydrogen. Maybe I should do it again with extra high effort?
👾HD-2D style house built with Gemini Hydrogen High. I really liked how Hydrogen came up with the idea of adding its own testing panel and used it to click through and test different effects during development. Something's a bit off with the lighting, but overall the result is pretty cool! On Google AI Pro plan it consumed 4% of 5h limit and 2% weekly which is really good. Done in 8 minutes.
1
111
New open weights Chinese model? I’ll take this!
A new open-weights DeepSeek V4.1-Flash competitor just appeared. It fits on 2x DGX-Sparks and can ... operate scientific software, check its own work, and FIX its mistakes. It uses the same architecture as GLM-5.3-Flash. This is Sage-S2.1-Flash from Shanghai AI Lab. Here are the model's stats ... 🧠 320B total / 18B active 📚 ~1M token context 👁️ Multimodal 🔓 MIT license 💾 BF16 + FP8 weights downloadable Shanghai AI Lab demonstrated it working with 12,487 mixed medical images. Example application, the model 👉 Identified 139 relevant CT slices 👉 Attempted a reconstruction in 3D Slicer 👉 Detected a slice-ordering inconsistency 👉 Corrected the error 👉 Reconstructed the skeleton in 3D It didn't just describe the images. Also, it operated scientific software and used verification feedback to correct its work. And Sage supports 10+ types of scientific inputs 🧬 DNA / RNA / protein sequences 🧪 Molecular graphs 💎 Crystal structures 📈 Spectra 🧠 EEG / ECG signals Now look at these developer-reported benchies. MLE-bench Lite: 🔥 Sage S2.1 Flash → 90.3 🐋 DeepSeek V4.1 Flash → 89.2 🟠 Claude Opus 5 → 88.0 ⚫ GLM-5.3-Flash → 84.3 🟣 Qwen3.8-Max → 82.2 🟢 GPT-5.6 Sol → 80.7 And compared with GLM-5.3-Flash: OfficeQA Pro 64.7 → 72.2 Harvey LAB 20.0 → 25.0 SpreadsheetBench 2 32.4 → 37.1 Terminal-Bench 2.1 84.3 → 87.9 Not a clean sweep across every benchmark, but these are interesting results for a scientific agent. ⚖️The weights are already here. 💾 BF16 → ~645GB 💾 Official FP8 → ~330GB Raw weight math: 4-bit → ~160GB 3-bit → ~120GB 2-bit → ~80GB Remember all the work Localmaxxers have done getting GLM-5.3-Flash running on affordable hardware? So who's going to quantize it for a 192GB Gorgon Halo? Or two DGX Sparks? 👀 Imagine running your own scientific research agent that doesn't just generate answers. It executes experiments, checks results, and corrects its mistakes. Link in reply
4
432
👾HD-2D style house built with Gemini Hydrogen High. I really liked how Hydrogen came up with the idea of adding its own testing panel and used it to click through and test different effects during development. Something's a bit off with the lighting, but overall the result is pretty cool! On Google AI Pro plan it consumed 4% of 5h limit and 2% weekly which is really good. Done in 8 minutes.
1
1
6
284
HD-2D style house built with Qwen 4 max. Not bad, but not great. I was expecting a bit more.
1
1
11
455
If this is true it’s huge. 😳
我去!! 原来 GPT 6.2 Sol 真的把 opus 5.5 蒸馏了 日常测试 GPT 6.2 Sol 灰度,去验指纹发现验证出来的是 Claude Opus 5.5。 怪不得之前流出的灰度效果和 Opus 那么像🤣🤣
1
3
439
Qwen 4 in action. 👇
Qwen 4 max is here, I've done A LOT of tests with it, many I never posted. The video contains many of the demos I ran over the week. The one thing I can say about Qwen 4 is that it's very token efficient. I'm cautiously optimisic about Flash and 27b.
1
7
244
Now we have to wait for 4 27b. ✌️
📢 Meet Qwen4-Max, your new AI teammate for work, creativity, and discovery! Qwen4-Max is built on a new MoE architecture with native vision and a 1M-token context window. Highlights: - Outperforms Qwen3.8-Max with up to ~50% fewer tokens, while average cost per task drops over 60% in our internal evaluation. - Advances real-world productivity through greater proficiency with professional tools, stronger long-horizon execution, and higher-quality deliverables. - Handles real tasks, including 3D modeling, game development, manufacturing, finance, law, healthcare, and frontier scientific discovery. - $1.5/M tokens for input, $4.5/M tokens for output, and $0.075/M tokens for implicit caching — all lower than Qwen3.8-Max, with implicit caching down by 70%. The API is live on QwenCloud. You can also try Qwen4-Max on Qwen Studio, QwenWork and Qoder. Start building today! 🚀 - Blog: qwen.ai/blog?id=qwen4-max - API: qwencloud.com/models/qwen4-m… - Token Plan: qwencloud.com/pricing/token-… More to come! The open weights of smaller Qwen4 models will be released. Stay tuned. 👀
1
5
269
Qwen 3.8 Flash Next. I can't believe this local model created the spec and built such a nice pixel-art React app! One prompt, few minutes.
1
2
210
‼️Anthropic updated its usage policy yesterday. What changed for Hermes and OpenClaw users? ✅ Official Claude Code CLI - no changes. ⚠️ DirectSDK with a local proxy - more regulatory uncertainty. ⛔ Direct OAuth - still prohibited, no changes. The new rules mainly target unauthorized proxies. Codex remains more open to third-party agents.
1
2
230
Friendly reminder. We will miss you.
For my first post, I’m sharing a letter @NVIDIA signed on why open models matter. AI will transform every industry, power every company, and be built by every country. Open models strengthen safety and cybersecurity, accelerate innovation and diffusion, and enable sovereignty. The world needs both frontier closed models and frontier open models. images.nvidia.com/pdf/Open-W…
102