Alibaba open-sources Qwen3.8-2.4T-A95B MoE model with 2.4T total parameters
Alibaba has open-sourced Qwen3.8-2.4T-A95B, its largest and most capable open-source model, a text-only MoE with 2.4 trillion total parameters and 95 billion active parameters. The FlagOS community announced Day-0 adaptation across nine AI chip platforms, offering BF16, FP8, and INT8 versions with unified open-source stack support. The model supports 262,144-token native context, extendable to 1,010,000, and includes reasoning_effort and preserve_thinking settings.
Coverage timeline
Hacker NewsPhilpax
# Set the following accordingly export OPENAI_BASE_URL='your-base-url' export OPENAI_API_KEY='your-api-key' ``` ##### Text-Only Input ``` from openai import OpenAI # Configured by environment variables client = OpenAI() messages = [{"role": "user", "content": "Write a Python function to merge two sorted linked lists."}] completion = client.chat.completions.create( model="Qwen/Qwen3.8-2.4T-A95B", messages=messages, extra_body={ "chat_template_kwargs": { "enable_thinking": True, # on by default; should not be turned off "preserve_thinking": True, # on by default }, }, reasoning_effort="xhigh", # xhigh by default; supported levels are xhigh, medium, and low stream=True, stream_options={"include_usage": True}, ) reasoning_content = "" answer_content = "" is_answering = False print("\n" + "=" * 20 + "Reasoning" + "=" * 20 + "\n") for chunk in completion: if not chunk.choices: print("\nUsage:") print(chunk.usage) continue delta = chunk.choices[0].delta if hasattr(delta, "reasoning_content")
# Set the following accordingly export OPENAI_BASE_URL='your-base-url' export OPENAI_API_KEY='your-api-key' ``` ##### Text-Only Input ``` from openai import OpenAI # Configured by environment variables client = OpenAI() messages = [{"role": "user", "content": "Write a Python function to merge two sorted linked lists."}] completion = client.chat.completions.create( model="Qwen/Qwen3.8-2.4T-A95B-FP8", messages=messages, extra_body={ "chat_template_kwargs": { "enable_thinking": True, # on by default; should not be turned off "preserve_thinking": True, # on by default }, }, reasoning_effort="xhigh", # xhigh by default; supported levels are xhigh, medium, and low stream=True, stream_options={"include_usage": True}, ) reasoning_content = "" answer_content = "" is_answering = False print("\n" + "=" * 20 + "Reasoning" + "=" * 20 + "\n") for chunk in completion: if not chunk.choices: print("\nUsage:") print(chunk.usage) continue delta = chunk.choices[0].delta if hasattr(delta, "reasoning_conten
机器之心新闻资讯
阿里巴巴开源超大规模MoE模型Qwen3.8-2.4T-A95B,众智FlagOS社区同步完成Day0多芯片适配。Qwen3.8-2.4T-A95B已在平头哥、英伟达、摩尔线程、华为昇腾、沐曦、昆仑芯、海光、清微智能、隧原等9家AI芯片上完成基于FlagOS统一开源技术栈的多芯适配、精度对齐与部署验证,针对不同芯片情况提供包括BF16FP8INT8等多种精度的版本,开发者可直接获取对应芯片的开箱即用方案。 从今年2月首次开展MiniCPM4.5-o模型的多芯片Day0适配,到今天实现Qwen3.8-2.4T模型在9款芯片上的Day0适配,FlagOS已累计完成来自7大头部模型团队、12款主流开源模型、覆盖多达10款芯片的跨芯Day0适配,向AI芯片与大模型产业验证了:基于统一、开放的系统软件栈,实现前沿模型"—次开发、多芯快速适配"已具备规模化落地能力。 本次发布的Qwen3.8-2.4T-A95B是阿里巴巴开源模型系列中规模最大、能力最强的一代,首次将Qwen-Max级别的模型开放出来。主要特性: 超大规模MoE架构,总参数2.4T,激活参数95B,纯文本模型 编程、专业工作、科研、长程智能体任务能力显著提升;支持reasoning_effort调节推理深度，preserve_thinking保留历史推理 原生262,144 tokens,可扩展至1,010,000;提供BF16/FP8权重,兼容vLLM、SGLang、TokenSpeed 在Coding Agent相关基准测试中,Qwen3.8-Max相比Qwen3.7-Max有明显提升:Terminal Bench 2.1从74.5升至86.6,SWE-bench Pro从60.6升至67.7,接近Opus 4.8(69.2),PaperBench从64.8升至93.0,并超过Opus 4.8(80.3)。 本次Qwen3.8-2.4T模型的适配,虽然相比Qwen3.5/Qwen3.6模型主体结构变化有限,但最重要的挑战是模型参数规模比Qwen3.5-397B扩大了6倍。为了让当下已经部署的主流AI芯片能跑起来，FlagOS需要解决因参数规模巨大带来的一系列系统优化问题。为了解决超大规模参数问题,本次FlagOS技术栈新增了面向多款AI芯片的统INT8量化支持,通过FlagOS-Compressor多芯片模型
