Qwen3.8: Alibaba's 2.4T MoE Model, Qwen3.8-Max API, Open Weights, Limits and Pricing

What Qwen3.8 is

Qwen3.8 is a model family from Alibaba's Qwen team, released on 3 August 2026. Its flagship, Qwen3.8-Max, is offered on Qwen Cloud under the model id qwen3.8-max, and Alibaba describes it as a 2.4-trillion-parameter mixture-of-experts model aimed at coding and professional work. Alibaba states that it can code and deliver complete projects over ten or more days, handle specialised tasks in legal, financial and design work in one conversation, and use native visual understanding across planning, execution and verification, including long documents and long videos. On LMArena's text leaderboard Qwen3.8 has 1484 Elo (rank 25, max effort).

Open weights

Alibaba published the base model on Hugging Face as Qwen/Qwen3.8-2.4T-A95B, with an FP8 version, under a licence named qwen3.8-max; it is not Apache 2.0. The model card lists 2.4T total and 95B activated parameters, 92 layers, and 512 experts with 10 routed and 1 shared expert active per token. The context length is 262,144 tokens natively, extensible to 1,010,000. The open model is text-only and always thinks: multimodal input is not supported and thinking cannot be switched off. Qwen3.8-Max is the official hosted version built on it, adding vision input, a non-thinking mode, a 1M context by default and built-in tools. The smaller Qwen3.8-27B is a separate release under Apache 2.0.

Qwen3.8-Max on Qwen Cloud

The hosted model accepts text, images and video and returns text. Qwen Cloud lists a context of 1M tokens, a maximum input of 991K tokens, or 983K with thinking, a maximum output of 131K tokens, a maximum reasoning length of 262K tokens, and a limit of 2,000,000 tokens per minute. Supported features include partial mode for continuing from a prefix, function calling, context caching, structured outputs, batch processing, web search and fine-tuning. Built-in tools include a code interpreter, web extractor, web search and text-to-image and image-to-image search through the Responses API, PDF parsing through the Completions API, and web fetch and web search through the Anthropic-compatible API. Reasoning depth is set with reasoning_effort.

Pricing

On Qwen Cloud, Qwen3.8-Max costs $2 per million input tokens and $6 per million output tokens. Input served from the implicit cache costs $0.25 per million tokens. Through batch files, input costs $1 and output $3 per million tokens. Creating an explicit cache costs $2.50 per million tokens and reading from it $0.17. The open weights carry no usage fee from Alibaba, so the cost of running Qwen3.8-2.4T-A95B depends on the hardware used to host it.

Benchmarks, and who ran them

The Hugging Face model card reports results on coding and agent benchmarks such as Terminal Bench 2.1, SWE-bench Pro, DeepSWE 1.1, NL2Repo-Bench, FrontierSWE and SkillsBench, alongside several in-house Qwen benchmarks. Alibaba notes that it ran many of these with the Claude Code harness and its own settings, that some competitor scores are the best published figures across different harnesses, and that the SkillsBench results come from its own testing. The scores are therefore Alibaba's own and are not directly comparable with figures run elsewhere.

Official Resources