Qwen3.8
A large language model from Alibaba based on a Mixture-of-Experts architecture with a total of approximately 2.4 trillion parameters. It supports continuous OTA updates and achieves state-of-the-art performance on multi-agent tasks including full-stack development and data analysis.
Timeline (last 90 days)
Without prefilling, this abnormal behavior is rare, but it can be reliably reproduced with a specific prompt combined with prefilling.
Developers discovered a "strange links" anomaly when running Qwen3.8 locally.
The author of that PR had previously been involved in work related to Qwen3.8.
Qwen3.8 thinks first, and a max_tokens of 64 may be exhausted before the answer appears; setting chat_template_kwargs = {"enable_thinking": false} disables thinking and returns only the answer.
On the MCP-Atlas (tool calling) benchmark, Muse Spark Open scores 89.2 and GLM 5.3 scores 84.2.
Qwen3.5, Qwen3.6, and Qwen3.8 gained faster prefill and speculative decoding support on MLX.
NVIDIA Nemotron 3、Qwen3.5 至 Qwen3.8、Kimi-K3 以及 GLM-5.3-Flash 等模型均采用了混合架构范式。
The recommended hybrid workflow is to use large models (such as ChatGPT) for planning and solution review, and local models for specific coding execution, so as to balance data privacy and cost control
Qwen3.8 is an open-source model with 27 billion parameters
The most critical hardware metric for running LLMs such as Qwen3.8 is GPU VRAM — the more the better
5 more timeline events
All Facts (14)
Without prefilling, this abnormal behavior is rare, but it can be reliably reproduced with a specific prompt combined with prefilling.
50%UnverifiedThe author of that PR had previously been involved in work related to Qwen3.8.
50%UnverifiedQwen3.8 thinks first, and a max_tokens of 64 may be exhausted before the answer appears; setting chat_template_kwargs = {"enable_thinking": false} disables thinking and returns only the answer.
50%UnverifiedOn the MCP-Atlas (tool calling) benchmark, Muse Spark Open scores 89.2 and GLM 5.3 scores 84.2.
50%UnverifiedQwen3.5, Qwen3.6, and Qwen3.8 gained faster prefill and speculative decoding support on MLX.
50%UnverifiedNVIDIA Nemotron 3、Qwen3.5 至 Qwen3.8、Kimi-K3 以及 GLM-5.3-Flash 等模型均采用了混合架构范式。
50%UnverifiedQwen3.8 is an open-source model with 27 billion parameters
50%UnverifiedThe most critical hardware metric for running LLMs such as Qwen3.8 is GPU VRAM — the more the better
50%UnverifiedThe recommended hybrid workflow is to use large models (such as ChatGPT) for planning and solution review, and local models for specific coding execution, so as to balance data privacy and cost control
50%UnverifiedQwen3.8 is not suitable for complex architectural design and project planning tasks
50%UnverifiedThe full-stack news aggregator implementation plan generated by Qwen3.8 was cross-validated by GPT 5.6, which identified 6 critical issues
50%UnverifiedQwen3.8 successfully completed in one shot, without any planned prompting, a Pomodoro timer application featuring multiple states, browser notifications, data analysis, and animations
50%UnverifiedDeepSeek V4 Pro的参数规模远小于Kimi K3或Qwen3.8这类模型
50%UnverifiedDevelopers discovered a "strange links" anomaly when running Qwen3.8 locally.
50%