XNews

Open-source & research model releases (Qwen, Meta CWM)

Open-source & research model releases (Qwen, Meta CWM)

Open-source and research model releases advanced across vendors. - Alibaba announced Qwen3-Next-80B FP8 variants and Qwen3-VL models achieving top OpenRouter market share for image tasks - Qwen3-VL-235B claimed 48% OpenRouter share for image processing - Qwen3-Omni claims end-to-end omni-modal unification and open-sourced multiple variants (instruction, thinking, captioner) - Meta FAIR released Code World Model (CWM), a 32B research model with competitive pass@1 coding and math results (SWE-bench Verified 65.8%, LiveCodeBench 68.6%, Math-500 96.6%, AIME 76.0%) - Qwen and Meta releases aim to provide open weights and developer tooling for research and production experimentation

Key insights

Alibaba advances Qwen3 series with FP8 Next and strong image model share

Alibaba rolled out Qwen3-Next-80B FP8 variants (Instruct and Thinking) to accelerate inference and efficiency. Qwen3-VL-235B and Qwen3-VL models reported large image-processing adoption on OpenRouter, with Qwen3-VL-235B cited at ~48% market share for image tasks. Alibaba also open-sourced Qwen3-Omni instruction, thinking and captioner variants to support multimodal and AV workloads.

Meta publishes Code World Model (CWM) 32B for research in code and reasoning

Meta FAIR released a 32B Code World Model (CWM) as an open-weight research model targeted at code and reasoning tasks. Reported evals include SWE-bench Verified 65.8%, LiveCodeBench 68.6%, Math-500 96.6% and AIME 76.0%, and Meta encouraged community research using the open weights.

Source posts