金のニワトリ (@gosrum)

Gemma-4를 thinking 없이 실행해 본 결과 성능이 떨어졌고, Qwen3.5와 달리 ts-bench 점수가 낮아졌다고 합니다. 성능을 중시한다면 Gemma-4는 기본값인 thinking 모드로 사용하는 것이 좋다는 실전 평가입니다.

https://x.com/gosrum/status/2039947867662799341

#gemma4 #qwen3.5 #benchmark #thinking #llm

New update for the slides of my talk "Run LLMs Locally": WebGPU

Now models can run completely inside the browser using Transformers.js, Vulkan and WebGPU (slower than llama.cpp, but already usable).

https://codeberg.org/thbley/talks/raw/branch/main/Run_LLMs_Locally_2026_ThomasBley.pdf

#ai #llm #llamacpp #stablediffusion #gptoss #qwen3 #glm #localai #webgpu

don't expect llm generated code to be correct ↓

Qwen3.6-Plus: Towards Real World Agents

https://qwen.ai/blog?id=qwen3.6

#HackerNews #Qwen3.6Plus #RealWorldAgents #AI #Technology #Innovation

버려지던 신호를 학습으로, OpenClaw-RL이 AI 훈련을 바꾸는 방법

https://aisparkup.com/posts/10208

Using #ollama with #Qwen3 (30b) to finally sort through the ~10K unread emails.

After it's done, I'll start reading them as they come, pinky swear! (I probably won't though, I know myself)

#ai #UsefulAI

New update for the slides of my talk "Run LLMs Locally":

Now including music generation with ACE-Step and OCR using LightOnOCR.

https://codeberg.org/thbley/talks/raw/branch/main/Run_LLMs_Locally_2026_ThomasBley.pdf

#ai #llm #llamacpp #stablediffusion #gptoss #qwen3 #glm #localai

Google Turbo Quant running Locally in Atomic Chat

MacBook Air M4 16 GB
Model: QWEN3.5-9B
Context window: 50000

Summarising 20000 words in just seconds..
You can do 3x larger context window, processing 3x faster than before!

https://x.com/atomic_chat_hq/status/2037650438024007760

#qwen3.59b #turboquant #atomicchat #macbookairm4 #contextwindow

Ivan Fioravanti ᯅ (@ivanfioravanti)

vllm-metal에서 PR 작업을 통해 M3 Ultra에서 최대 32K 컨텍스트로 Qwen3-0.6B 테스트를 실행한 내용이 공유됐다. mlx-lm이 여전히 더 낫다고 언급하며, TTFT는 적절한 지표가 아니고 개선 작업이 진행 중이라고 밝혔다. 히트맵과 Prefill/Decode 차트도 업데이트됐다.

https://x.com/ivanfioravanti/status/2036156678035419263

#vllm #mlxlm #qwen3 #llm #opensource