Qwen/Qwen3.5-4B

Timeline

1. Oct 19, 2026 · Opinion / analysis · 1 source
   Kapitsa: open-source Apple Silicon optimized Metal engine built for Qwen and Nemotron. See benchmarks of Qwen 27B, Nemotron + 8 other models on an M1 Pro (16GB), M4 Pro (48GB) and M5 Max (128GB)
   Kapitsa runs only two model families: Qwen (3.5 4B, 3.8 27B, 3.6 35B-A3B, and 3.8 Flash-Next on 128 GB Macs) and NVIDIA Nemotron (3.5 Lightning, Bolt and Super-VL, plus Dory 4B (this is new, we didnt test Dory too much yet). focusing on only these models allows us customize the Metal kernels for each one on Apple Silicon
2. Oct 15, 2026 · Opinion / analysis · 1 source
   QUASAR 4-bit QAT checkpoints for Qwen3.5, Gemma 4 and Muse Glimmer: closer to BF16 than Google’s and NVIDIA’s 4-bit releases
   Two months ago we released a fully quantized NVFP4 Qwen3.8-27B trained with QUASAR.
3. Oct 15, 2026 · Research paper · 2 sources
   MJev-Spatial: System One Spatial Intelligence via Native Attention Matching
   In light of this, we introduce MJev-Spatial, a System 1 multimodal decision model that tackles all three types of spatial understanding tasks within a unified classification framework.
4. Oct 15, 2026 · Research paper · 2 sources
   Mosaic: Multi-Session Agentic Reinforcement Learning
   Therefore, we introduce Multi-Session Agentic Reinforcement Learning, a more general setting that supports an arbitrary number of sessions within a trajectory.
5. Oct 14, 2026 · Model release · 1 source
   microsoft/UserRepGen-4B
   microsoft published the model UserRepGen-4B on Hugging Face.
6. Oct 14, 2026 · Research paper · 1 source
   Selective Reasoning for Typed Decision Models: When to Start, Read, and Stop
   Typed decision models need an interface that can use additional reasoning while retaining candidate outputs.
7. Oct 12, 2026 · Research paper · 1 source
   Releasing the Inner Jev: A Simple Recipe for Open Decision Models
   We show that existing open instruction-tuned language models can achieve comparable performance without any additional training, by reading out answers from letters (A, B, C, etc.) in the model's output logits at the last token position.
8. Oct 12, 2026 · Opinion / analysis · 1 source
   U-Space: Uncovering When and Why Uncertainty Arises in Language Models [R][D]
   If you build on open-weight reasoning models, you have probably wanted a "how sure is the model about this answer?" signal that does not cost 5x sampling, a judge model, or a labeled dataset.
9. Oct 12, 2026 · Research paper · 1 source
   Label-Centric Knowledge Retrieval for Zero-Shot Text Classification
   We propose SLKR (Semantic Label Knowledge Retrieval), a label-centric retrieval-augmented generation framework that separates offline knowledge construction from lightweight online inference.
10. Oct 11, 2026 · Opinion / analysis · 1 source
    Local Voice Assistant based on Qwen3.5 4B with skills/toolcalling. No dedicated GPU. Local STT+TTS
11. Oct 10, 2026 · Opinion / analysis · 1 source
    [AINews] TypeSafe/Jev at >$100M ARR, $7.5B valuation 3 weeks after launch
    As you can see in the AINews X recap section below, everyone on earth has cloned the Jev API, but only one company can ever create the category.
12. Oct 9, 2026 · Research paper · 1 source
    BoundaryEvo: Data-Model Co-Evolution at the Capability Boundary for Long-Chain Reasoning Distillation
    Distilling long-chain reasoning trajectories from large teacher models provides an effective way to improve the reasoning ability of small language models.
13. Oct 9, 2026 · Research paper · 1 source
    ClawLoop: Self-Improving Agents via Harness-Native Training Loop
    To address these limitations, we propose ClawLoop, a framework that enables self-improving without any external teacher model and with minimal prior dependencies.
14. Oct 8, 2026 · Research paper · 2 sources
    One Model to Loop Them All: Elastic Latent Recurrence for Language Models
    We present Elastic Latent Recurrence (ELR), a recipe for training looped language models without this tax.
15. Oct 6, 2026 · Research paper · 2 sources
    VisionWeave: Weaving Elastic Visual Representations as a Native Capability of MLLMs
    Multimodal large language models have become the dominant paradigm for visual understanding, but incur substantial costs by encoding inputs into dense, fixed-size patch tokens.
16. Oct 6, 2026 · Research paper · 2 sources
    FC-SWE: Failure-Conditioned RL for Long-Horizon Software Engineering Agents
    We introduce FC-SWE, a failure-conditioned RL framework that incorporates recovery attempts into policy training.
17. Sep 30, 2026 · Opinion / analysis · 1 source
    SCLATE: A Substrate for Continual-Learning Agent Training and Evaluation
    SCLATE: A Substrate for Continual-Learning Agent Training and Evaluation

Often appears with