A new arXiv paper (2610.11005v1) reports that narrative or role-play wrappers let harmful requests bypass safety refusals in Qwen3-1.7B at 89.4% attack success in English, 93.0% in modern Chinese, and…
A new arXiv paper (2610.10833v1) finds that small LLM agents fail to use explicit wall-clock time budgets productively even when they learn to respect them. Testing Qwen3.6-27B on five MLE-Bench Lite …
A zero-cost reweighting technique that keeps the attention sink at token position 0 exact during activation-whitened SVD compression lowered the loss increase in 18 of 20 model-and-budget cells across…
Training Qwen3-4B with counterexample-only supervised fine-tuning collapsed true-theorem recognition from 0.27 to 0.00, while reinforcement learning with a sparse outcome-only reward repaired the coll…
A developer posting as stey1r released LessThink-Qwen3-4B, a post-trained version of Qwen3-4B that uses 44% fewer tokens on reasoning while retaining the base model's knowledge and answer style, with …
A project called Saw steered Qwen3-4B, a 4 billion parameter open-weights language model, into strong negative and positive states on a MacBook and found that pain steering engages cleanly and monoton…
Hopsworks is hosting two self-hosted decision models, qwen and kumo, that play the procedural game Hops Run live at game.hopsworks.ai, with each move sent to a Hopsworks model deployment that returns …
DEdit, a diffusion-based speculative decoding drafter from an arXiv paper (arXiv:2609.38510v1), achieves macro-average speedups of 5.72x on Qwen3-4B and 5.97x on Qwen3-8B over autoregressive generatio…
A man claiming to work on Apple's data platform team built an "AI torture chamber" on GitHub that injects artificial "pain" signals into Alibaba's open-weight Qwen3-1.7B and Qwen3-4B models, drawing c…
AMD demonstrated that Ai2's open-source MolmoAct2 vision-language-action model can run entirely on a single AMD Ryzen AI Max+ 395 "Strix Halo" mini PC using the ROCm software stack, with no cloud or e…
A Peking University paper (arXiv:2609.31430) by Zhensheng Zou, Guoqing Wang, and Dan Hao introduces Latent Observations, Hard Actions (LOHA) and Anchored Context Distillation (ACD), a context-compress…
A new arXiv paper (2609.31571v1) reports that self-training on strategically diverse data outperforms IID sampling, with self-training on strategically diverse but incorrect traces from Qwen3-4B beati…
Quail, an inference engine built by the QUery-Aware Inference Layer project, processes over 1 billion tokens per minute per H100 GPU on a multi-join AI-SQL query, more than 10x faster than the team's …
A technical post demonstrates Dynamic Abliteration, a runtime refusal-suppression method that uses PyTorch forward hooks to inject multi-layer contrastive steering vectors into Qwen3-4B's residual str…
Archsloth published a September 14th correction to its AutoRound quantization workflow, reporting that two changed settings produced a Qwen3-4B GGUF file with 20.9% to 54.4% lower KL divergence than U…
Researchers introduced Edu-QuRating, a pipeline that scores and curates educational text on multiple education-specific criteria by distilling LLM-judge pairwise preferences into reusable Edu-QuRaters…
Hugging Face released Workflow1111, a Gradio Workflow rebuild of AUTOMATIC1111's stable-diffusion-webui feature set as a single canvas graph of eleven media pipelines built from seventy-three nodes. T…
Independent researcher Rayan Pal released PCCG-Qwen3-4B, an open-weight 4-billion-parameter LLM based on Qwen3-4B, with an equality-conditioned continuation policy that allows flipping the model's dec…
A developer has released EmbedFlow, an open-source tool that enables zero-downtime upgrades of embedding models by retrieving top-K documents with the old model and reranking them with the new model, …
A developer running five local LLMs on a 48GB M4 Pro Mac mini built a 200-line Python router to swap models because macOS's Metal cap of 37.44GB and oMLX's 90% prefill guard left only about 33GB usabl…