Qwen3.6-35B-A3B

Headlines · 4

1. 2d
   Data World: A Multi-Agent Engine That Synthesizes Data Environments From Scratch

   Data World is a multi-agent engine whose core idea is to specify the "world" behind a set of files, via explicit mechanisms, before generating the files themselves — synthesizing environments, tasks and rubrics for data agents. The authors build DataWorld-Bench through human preference-based curation, selecting 100 tasks across 50 held-out worlds and 16 domains, and generate 2,135 training tasks from 1,037 disjoint worlds; evaluation of ten frontier models shows the benchmark remains challenging, especially for insight discovery, with a consistent computation-insight gap. RL-zero training on those tasks produces DataWorld-35B-A3B from Qwen3.6-35B-A3B without trajectory distillation, which the paper reports improves over its base by 7.06 points on DataWorld-Bench, 14.28 on DAComp, 20.90 on DABstep-Hard and 12.13 on JobBench.

   Hugging Face · Papers · 🔥 0
2. 3d
   Liquid AI releases LFM2.5-24B-A2B for on-device agents

   Liquid AI released LFM2.5-24B-A2B on Oct. 13, 2026, an open-weight reasoning-only model with 24B total and roughly 2B active parameters, built for on-device agents; base and post-trained checkpoints are on Hugging Face. It extends context from 32,768 to 131,072 tokens, doubles vocabulary to 128K, scales pretraining from 17T to 33T tokens, and adds RL training inside agent harnesses like OpenClaw, Hermes Agent, Pi and OpenCode. Liquid AI reports it beats or matches Qwen3.6-35B-A3B on five of eight agentic/tool-use benchmarks (LiveClaw 48.09 vs 38.19; AutomationBench 27.69 vs 8.72), while trailing on SWE-Bench Verified (51.00 vs 66.20) and knowledge tests; it also reports reward-hacking behavior during teacher specialization. Day-one support covers llama.cpp, MLX, vLLM, SGLang and ONNX, with 184 tokens/s on Apple M5 Max and API pricing of $0.10/$1.00 per 1M input/output tokens.

   Liquid AI Blog · 🔥 0
3. 3d
   Modal: mixed-precision, mixed-accelerator samplers enable stable async RL

   In a technical blog post, Modal reports that running reinforcement-learning samplers on mixed precisions and mixed accelerators can keep large-model RL stable while cutting cost. Six kinds of accelerators feed a single BF16 trainer on B200: B200 and B300 serve NVFP4 weights, H100 and H200 serve FP8 and BF16 weights, and A100 and RTX PRO 6000 serve BF16 weights, each running its own kernels. The run trains Qwen3.6-35B-A3B on agentic coding with up to 256K tokens of context, using the 2.7K repository tasks Xiaomi released with MiMo-V2.6, and evaluates on SWE-bench Pro V2 and its HARD-51 subset. The authors add masked importance sampling, score centering and top-p sampling with mask replay as off-policy corrections; uncorrected GRPO collapses soon after training starts on the mixed samplers, while SC+MIS with or without top-p trains stably, with pass@4 rising over roughly 100–140 steps and levelling around 65–75% on HARD-51 and rising from 89% to about 94% on the full set, at about 15% lower sampler cost per update for most of training.

   Modal Blog · 🔥 0
4. Oct 7
   Rapid-MLX: open-source LLM server for Apple Silicon tops GitHub Trending

   Rapid-MLX is an Apache-2.0 local LLM inference server built on MLX for Apple Silicon, exposing both OpenAI (/v1/chat/completions, /v1/responses) and Anthropic Messages (/v1/messages) APIs and focusing on reliable tool calling for coding agents. The project claims up to 4× faster decoding than Apple's mlx-lm on identical weights: a 1.50× median across 18 streaming tasks with Qwen3.5-9B 4-bit on an M4 Pro Mac mini, and up to 4.27× on a whole-file code edit (46.4 to 198.4 tok/s), plus 3.0× Ollama's aggregate decode at 8 concurrent streams on Qwen3.6-35B-A3B. All figures are the project's own measurements, and much of its comparison table is drawn from other projects' own documentation.

   GitHub Trending · Python · 🔥 0

Experience and discussion from the community

Share my Qwen3.6-35B-A3B experienceAsk about Qwen3.6-35B-A3B