Qwen3.8-Max
Alibaba (Qwen) · Multimodal model
Qwen3.8-Max is Alibaba's flagship model, announced August 3, 2026, a 2.4-trillion-parameter mixture-of-experts model with a 1M-token context and image and video understanding. Its base weights were published on August 12 as Qwen3.8-2.4T-A95B.
Current Open weights
ReleasedAug 3, 2026
Context window1M tokens
Input → outputtext, image, video → text
API priceqwen3.8-max (International/Singapore, 0 < tokens ≤ 1M): $2 / 1M input tokens, $6 / 1M output tokens
Consumer accessQwen Studio (chat.qwen.ai), described as free to use; also QwenWork
Best for
Frontier-level coding and long-horizon agent work, with the option to self-host the open base checkpoint.
Overview
Qwen3.8-Max is Alibaba's flagship model, announced August 3, 2026, a 2.4-trillion-parameter mixture-of-experts model with a 1M-token context and image and video understanding. Its base weights were published on August 12 as Qwen3.8-2.4T-A95B.
Strengths and limits
- 2.4T total / 95B active MoE, Alibaba's largest model
- Open weights released (Qwen3.8-2.4T-A95B), first open Qwen-Max-class model
- Alibaba reports top-5 Text Arena and 2nd Vision Arena placement
- Open checkpoint is text-only and always-thinking; vision and non-thinking are API-only
- Custom 'qwen3.8-max' license, not Apache 2.0; ~4.9 TB of weights to self-host
Published benchmark scores
SWE-bench Pro67.7
Source: huggingface.co
Terminal Bench 2.186.6
Source: huggingface.co
GPQA Diamond92.6
Source: huggingface.co
HLE43.6
Source: huggingface.co
Scores are as published by the source; settings (reasoning effort, tools, attempts) differ between labs, so compare with care. How to read benchmarks →