Qwen3-0.6B mixes 4-bit decode with selective 8-bit steps
An entropy gate at a 16% escalation rate scores 62.7% on 150 GSM8K problems, using 0.87 of all-high bytes.
Stories per day · 30 days
23 RSS
Relatedreinforcement-learning17llm-agents11alibaba9fine-tuning9gemma9agent-security8benchmarks8cohere8llama8open-weights8huggingface7benchmark6coding-agents6deepseek6inference6open-weight6quantization6agents5
303 stories, newest first · one card per story
Stories tagged #qwen
Sable proves bit-exact Qwen inference on one GPU
On one RTX PRO 6000, Sable proves Qwen2.5-32B in 92.5 seconds for 4,096 tokens, about 6.4 cents per request.
Study tests whether causal neuron groups are unique or transferable
A few neurons often flip two-hop answers, but effects weaken on other chains.
Matched compute reverses dense-reward gains over GRPO
Across 1,668 runs, a Monte-Carlo process reward trailed GRPO by 14.3 points at equal tokens on Qwen2.5-1.5B.
Persona generalization follows a shared line across Qwen and Llama
Fine-tuning only later layers keeps strong in-domain learning while reducing out-of-domain unsafe transfer.
Game-theoretic monitor cuts deceptive behavior in agent benchmarks
On Gemma3 and Qwen3, e-PoG beat the strongest baselines by 26.0% on deception mitigation and up to 39.8% on task success.
ShadowBench-MCP measures shadow invocations in MCP-style tool routing
On a Groq-hosted qwen3-32b router, efficiency-first routing had a 29.2% shadow-invocation rate versus 4.2% when policy-aware.
Comfy-Org republishes MiniMax-H3-Image weights for ComfyUI
The pack adds int8 and NVFP4 files plus Qwen encoders, under a license that restricts high-revenue hosted use.
MoE Routing Strategies: Top-K, Expert-Choice, Sigmoid Gating and Dynamic Routing Explained
The guide covers seven production MoE routing methods used in models like Mixtral and DeepSeek-V4, highlighting sigmoid gating and bias balancing as dominant approaches.
1
Nous Research fine-tunes Qwen3.8-27B to strip Chinese propaganda
Nous Research fine-tuned Alibaba's Qwen3.8-27B on 53 million tokens to drop propaganda from 22 percent to under 1 percent in Arete-27B while keeping capabilities intact.
1
Spatial Harness grounds VLM spatial reasoning in physics-engine states
On MindCube it raises Qwen3.5-27B from 52.70% to 71.14%, and to 98.25% when poses are annotated.
Open-weight models do not report injected pain states
Across 52,920 reports, excess pain claims stayed under a few points, while probes still recovered the steered state.
Weight-Space Programming enables localized neural edits with up to 29.45-point gains
16 trajectories on Qwen models reveal WSP's strong localized improvements but inconsistent generalization and reliability
RidgeFill speeds prefill with a training-free late-layer cache
A closed-form ridge map fills upper-layer cache; reported TTFT gains are 1.25–1.60x on tested Qwen setups.
arXiv studies training data for open models on ARC-AGI-2
Fine-tuning open models with specific rules and traces improves scores on ARC-AGI-2, with the best reaching 56.67.
Astra reads and edits other models' weights without full retraining
Researchers say changing 0.016 percent of Qwen3.5-0.8B weights raised MATH-500 accuracy from 7.9 to 48.1 percent.
1
Qwen 3.8 27B reads and edits model weights on a laptop
Simon Willison says a local Qwen 3.8 27B reproduced a paper example previously shown with GPT-6 Astra.
LLMs are not superhuman at reading and writing neural network weights
A recent paper claims frontier LLMs can work with neural network weights directly, but this is not superhuman.
1
Qwen3.5-2B Compiled into qwenGPU with 45 CUDA Layers
The small model now runs 45 layers natively on GPU using CUDA.
CCORE GLB-3D-Coder 4B model trending on Hugging Face
4B model converts text prompts to editable 3D assets exportable as GLB files for Godot or Three.js.
1
Vals study finds Chinese AI models politically mixed, not mere mouthpieces
Simplified Chinese prompts increased pro-government answers, and most models answered sensitive questions rather than refusing.
1
Team of 5 Qwen3.8-27B models beats Fable 5 on LiveCodeBench Hard
An ensemble of five Qwen3.8-27B models outperforms Fable 5 on the hard LiveCodeBench benchmark.
Apple Intelligence is coming to China with iOS 27.3
A November 4 beta adds earlier features and Alibaba’s Qwen, not the new Siri; the full release is set for December.
NES 6502 emulator encoded in Qwen3.5-4B weights
Drew Breunig’s nesformer can run Super Mario Bros inside Qwen, very slowly, using a method inspired by qwenputer.