event

Qwen Model Update

17 items · started 2026-11-06 · ongoing (last activity 2026-11-16)

  1. This is a browser I am building using AI Agents. Components: Boa for JS, Stylo for CSS, Blitz for HTML DOM.

  2. Abliteration, a fine-tuning-free weight edit, removes refusal from open-weight large language models (LLMs). The developer of Arbitrary-Rank Ablation (ARA), an optimisation-based variant, reports that direction-based abliteration, which ab…

  3. Qwen 4 is about to come out (maybe?? We don’t know Qwhen exactly 😁) and since it’s likely it’ll replace my model of choice, Qwen 3.8 Flash Next (or QFN for short), I felt this was a good moment to shed some light on a model that changed ev…

  4. I built local-model-mcp, a small MCP server that lets a Claude session hand light, read-only work to a model running locally under Ollama. I designed it iteratively with Claude Code, deciding at each step which functions to add next, and C…

  5. RUN 125B+ MoE ON CONSUMER GPUS Powered by Niko1221's dynamic sparse router offloading. Run server-grade frontier models like Qwen3.8-Flash-Next 125B on a single 12GB NVIDIA graphics card and 64GB system RAM—with zero cloud dependencies.

  6. Qwen3.8 27B Reasoning Trainer This is a Pure Go Abliteration, Reasoning Training, and Inference stack. It is implemented with a CPU backend and a Vulkan backend, which uses compute shaders and GPU kernels.

  7. Build More, Spend Less Token Plan for Individual is now available from $3/month. Be the first to try Qwen3.8-Max and enjoy advanced models with all modality in one plan.Coupon Rules One plan.

  8. Byte-level BPE fragments Chinese text (for Qwen3.8 on Chinese Wikipedia, 3.84 bytes per token on average; 64% of tokens are single characters) and hides spelling from the model. We ask whether an already-trained 27B subword model can be co…

  9. $2.98 and 11 minutes. Fable 5.2 working alone took $15.77 and 21 minutes on the same prompt.

  10. Yes, it is yet another fork. But i think it brings some interesting things to the table.

  11. Use balena etcher or whatever to mount to USB stick, boot. Test.

  12. Well I had a lot of resets on ChatGPT and all of those credits and I thought, well can Astra actually do it. We had a little looping and a lot of 'Wait.' Let me check the thing I just checked -- but it always got the job done.

  13. I need to make my custom harness much better at general research and other tasks (not coding) that run longer than half a minute. Currently running: Qwen3.8 Flash Next FP8/GLM 5.3 Flash W4A16 Relevant capabilities already in my harness: Th…

  14. ran the new ripple-65b-preview (q4_k_m, llama.cpp, 64 gb macbook pro, no cloud) through atomic agent on a 5-part python app graded by 50 hidden tests. 3 runs.

  15. A few weeks ago we posted Tessera-27B-B-CJK here, a Qwen3.8-27B coding agent that thinks in compressed CJK and uses fewer thinking tokens per agent turn. Thanks for all the feedback on that one.

  16. Tessera-27B-B-CJK-RL Tessera-27B-B-CJK-RL is a coding-agent model built on Qwen3.8-27B. It thinks in compressed CJK, a shorthand that mixes English, Chinese words and symbols, and it thinks less per agent turn than the base model.

  17. VibeSys on Trainium: Agents Build and Tune Qwen3.8 Serving Engines on AWS Trainium2 Keisuke Kamahori, Shihang Vic Li, Simon Peter, Baris Kasikci New ResearchInference & ServingAgents November 4, 2026 TL;DR: We ran VibeSys, our multi-agent…

← all threads