What it is
Qwen3.8-Max is Alibaba's flagship Qwen model, announced on 3 August 2026. Its weights are published, in BF16 and FP8, as Qwen3.8-2.4T-A95B: a mixture-of-experts model with 2.4 trillion parameters, 95 billion active per token. The open weights are text-only with a 262K-token native context (extendable to about 1M) and always think before answering.
The hosted Qwen3.8-Max on Qwen Cloud adds image and video input, a non-thinking mode, a 1M-token context, built-in tools and up to 131K output tokens. It costs $2 per million input tokens and $6 per million output tokens.
The weights use a custom Qwen licence. It allows commercial use, but model-hosting and AI-assistant businesses with over $50 million in yearly revenue need a separate licence, and very large products must show the model name. Smaller Qwen3.8 models, a 27B and the 180B Flash-Next, are also available.
Our review
Qwen3.8-Max is a frontier-class model at a mid-tier price, but the version most people can use is the hosted one. Qwen's table compares it with Claude Opus 4.8, Claude Fable 5 and GPT-5.6 Sol: it leads on PaperBench (93.0, reproducing research papers), IFBench (82.8 against 72.7 for GPT-5.6 Sol), HealthBench (60.2) and PRBench-Finance (58.3). It trails Fable 5 on most coding-agent tests, including SWE-bench Pro (67.7 against 80.0), and on Humanity's Last Exam (43.6 against 53.3).
The footnotes matter. Qwen ran its own model in Claude Code over ten attempts for Terminal-Bench 2.1 but took the best published score for rivals, re-ran SWE-bench Pro on a corrected task set, and several benchmarks, such as QwenSWEBench and CoWorkBench, are in-house. Outside evidence is thin: Mistral's 6 October Large 4 launch claims a higher Coding Agent Index score than Qwen3.8-Max without publishing Qwen's figure.
The open weights are a hedge, not a deployment plan. At 2.4 trillion parameters the FP8 checkpoint alone is roughly 2.4 TB, more than the 1.1 TB of memory in an eight-GPU H200 node by our arithmetic, so self-hosting means multi-node serving. The weights are also text-only and cannot turn thinking off; image and video input, non-thinking mode, the 1M default context and built-in tools exist only on Qwen Cloud. Hugging Face shows about 25,000 downloads of the full weights in the past month, against 2.7 million for Qwen3.8-Flash-Next.
Oddly, the flagship carries the looser licence. Its separate-licence rule for hosting and AI-assistant businesses applies only above $50 million in yearly revenue, while the smaller Flash-Next's applies at any size, so the model few can run is the one most companies are free to build on.
Where it earns its place
A research team at a pharmaceutical company
Reproduce methods from papers and draft analysis code
Call the hosted model with thinking on; Qwen reports 93.0 on PaperBench, ahead of the Claude and GPT models it tested. 50 million input and 10 million output tokens a month cost about $160 before cache savings.
A legal operations lead at a 200-lawyer firm
Produce contract summaries that follow strict house templates
Its IFBench score of 82.8 is the highest in Qwen's table, which matters when format rules are rigid. Qwen Cloud's model page does not state retention or training terms, so check the service agreement before sending client documents.
An infrastructure team at a regional sovereign-cloud provider
Offer a frontier-class open model to its customers
The weights need multi-node serving, roughly 2.4 TB in FP8 by our arithmetic. A model-as-a-service business above $50 million in yearly revenue needs a separate licence from Qwen before launch.
Before you commit
- Setup
- Hosted: a Qwen Cloud key and an OpenAI-compatible call to qwen3.8-max, with enable_thinking set as needed. Self-hosted: BF16 or FP8 weights served with vLLM, SGLang or TokenSpeed across several nodes.
- Integrations
- OpenAI-compatible Chat Completions and Responses, an Anthropic-compatible API and DashScope; function calling, structured outputs, fine-tuning and built-in web search, code interpreter and PDF parsing.
- Data and privacy
- Licence allows commercial use; model name must be shown above 100 million monthly users or $20 million monthly revenue; separate licence for model-as-a-service or AI work assistant businesses above $50 million yearly revenue; internal use is exempt. Qwen Cloud's model page does not state retention or training use.
- Cost at scale
- 10 million input and 2 million output tokens a day, a busy internal assistant, costs about $32 a day or roughly $960 a month at list price; explicit cache reads at $0.17 cut the cost of repeated context by over 90%.
- Limits
- Hosted: 1M context (991K input), 131K output, up to 262K reasoning tokens and 2 million tokens a minute. Open weights: text only, thinking always on, 262,144 tokens native and about 1 million with extension.
What's good
- Leads Qwen's table on PaperBench, IFBench and HealthBench
- Hosted API at $2 / $6 with 1M context and image and video input
- Open weights with commercial use allowed
- Licence restrictions start only above $50M yearly revenue
What's not
- Open weights are text-only and always think
- Self-hosting needs multi-node hardware
- Trails Claude Fable 5 on SWE-bench Pro and HLE
- Qwen Cloud's model page states no retention or training terms
Best for
Teams doing research, specification-heavy writing and multilingual work through an API who want an open-weights fallback.
Alternatives
Qwen3.8-Max: common questions
- Is Qwen3.8-Max worth it?
- An open flagship that follows instructions and reproduces research better than Claude or GPT on Qwen's numbers, though few will run the weights themselves. We rate it 4.2 out of 5.
- How much does Qwen3.8-Max cost?
- Open weights under the Qwen3.8-Max licence. Qwen Cloud API: $2 input / $6 output per 1M tokens; cache reads $0.17 (explicit) or $0.25 (implicit).
- What are the downsides of Qwen3.8-Max?
- Open weights are text-only and always think; self-hosting needs multi-node hardware; trails Claude Fable 5 on SWE-bench Pro and HLE; qwen Cloud's model page states no retention or training terms.
- Who is Qwen3.8-Max best for?
- Teams doing research, specification-heavy writing and multilingual work through an API who want an open-weights fallback.
- How much does Qwen3.8-Max cost for a team or in production?
- 10 million input and 2 million output tokens a day, a busy internal assistant, costs about $32 a day or roughly $960 a month at list price; explicit cache reads at $0.17 cut the cost of repeated context by over 90%.
- Is Qwen3.8-Max safe for company data?
- Licence allows commercial use; model name must be shown above 100 million monthly users or $20 million monthly revenue; separate licence for model-as-a-service or AI work assistant businesses above $50 million yearly revenue; internal use is exempt. Qwen Cloud's model page does not state retention or training use.
- What are the best Qwen3.8-Max alternatives?
- The closest alternatives we have reviewed are GLM-5.3, Kimi K3, DeepSeek-V4-Pro and GPT-6.1 Sol.
Answers checked Oct 23, 2026.
Make Qwen3.8-Max? Show this rating on your site
