Qwen3.5-4B

Headlines · 4

1. 2d
   Mosaic: multi-session agentic reinforcement learning

   A paper submitted to arXiv on 15 Oct 2026, "Mosaic: Multi-Session Agentic Reinforcement Learning," introduces multi-session agentic RL, a more general setting in which a trajectory may span an arbitrary number of sessions to match inference-time execution such as context compaction and subagent forking. The authors note that the common practice of broadcasting trajectory-level advantages to all sessions neither distinguishes session contributions nor guides session allocation, so they propose Mosaic, a critic-free hierarchical method that folds session-usage signals into trajectory-level advantage estimation and shapes each session's advantage by its estimated marginal contribution. Experiments on BrowseComp with Qwen3.5-4B show Mosaic outperforming broadcasting in both context compaction and subagent forking scenarios while using fewer sessions on average.

   Hugging Face · Papers · 🔥 0
2. 2d
   MJev-Spatial: spatial intelligence via native attention matching

   A paper on arXiv introduces MJev-Spatial, a "System 1" multimodal decision model that unifies three kinds of spatial understanding — multiple-choice, numerical prediction and 2D point prediction — as classification: candidates are scored directly with the backbone's pretrained query and key projections, adding no parameters, and each decision takes a single forward pass that generates no answer tokens. Built on Qwen3.5-4B and trained on 12.3M question-answer pairs converted from existing spatial datasets, the authors report the highest overall score (65.64%) among eight tested models, first place on seven of twelve benchmarks, and better results than its unmodified autoregressive base model on all twelve, while running 2.5x faster than the fastest autoregressive baseline and 5.6x faster than the base model. The paper also reports that joint supervision of coarse-to-fine stages plus length-aware global batching and cross-GPU load balancing raised training throughput by 30.8% over the original batching strategy.

   Hugging Face · Papers · 🔥 0
3. 4d
   AWS details llama.cpp decision-model serving on SageMaker AI with SemIf

   AWS published a walkthrough for serving decision models with llama.cpp on Amazon SageMaker AI using specialized Deep Learning Containers. The standard path deploys a GGUF model from Amazon S3 and maps InvokeEndpoint to llama-server's OpenAI-compatible chat completions; an optional SemIf child image (SemIf was formerly OpenJev) returns probabilities over caller-declared options by reading final-position logits instead of decoding an answer token. On 231 public JevBench tasks with Qwen3.5-4B Q4_K_M, a 32-thread CPU endpoint (ml.c6i.8xlarge) was correct on 179/231 (77.5%, p50 1.715s, p95 19.291s), while full GPU offload on ml.g6e.xlarge scored 180/231 (77.9%, p50 0.352s, p95 0.599s) — 4.9x faster at p50 and 32.2x at p95. AWS says the two images used different llama.cpp builds and disagreed on six hard tasks, so it calls the accuracy gap runtime variation rather than proof that GPU offload improves quality, and cites Jev, Kev and Strands Decider 2B as related decision-model work.

   AWS Machine Learning Blog · 🔥 0
4. Oct 9
   Unsloth releases free notebook to train decision models locally

   Unsloth published a free Colab notebook and companion guide showing how to train Qwen3.5-4B locally into a “decision model” that outputs decisions instead of text, using only 8GB of VRAM. Unsloth says its open-source repo can turn any model, such as Qwen3.8 or Gemma 4, into a decision model, and claims it raised Qwen3.5 0.8B’s aggregate accuracy from 20.7% to 74.3% across three decision benchmarks on just 4GB of VRAM (these figures are Unsloth’s own and the benchmark methodology was not disclosed). The walkthrough covers data preparation (state, questions, gold answers), training and serving, and likens the result to Jev.

   UnslothAI (X) · 🔥 0

Experience and discussion from the community

Share my Qwen3.5-4B experienceAsk about Qwen3.5-4B