Qwen3.8-Max-0902
Also known as: qwen3.8-max-2026-09-02
Qwen3.8-Max-0902 is Alibaba's September 2, 2026 upgrade to Qwen3.8-Max, also exposed as qwen3.8-max-2026-09-02. It retains 2.4T parameters and a 1M-token context window while adding post-training for coding and cowork tasks, with improvements to engineering-scale coding, long-horizon agents, multi-tool workflows, and native vision. QwenCloud pricing is $2 per million input tokens and $6 per million output tokens.
Data sourced primarily from official releases (GitHub, Hugging Face, papers), then benchmark leaderboards, then third-party evaluators. Learn about our data methodology
| Type | Condition | Input | Output |
|---|---|---|---|
| Text | - | $2.00/ 1M | $6.00/ 1M |
| Type | TTL | Write | Read |
|---|---|---|---|
| Text | - | $2.50/ 1M Cache = explicit | $0.170/ 1M Cache = explicit |
| Text | - | — | $0.250/ 1M Cache = implicit |
“—” means the modality is not billed in that direction, or the vendor has not published a price for it.
Qwen3.8-Max-0902 currently shows benchmark results led by NL2Repo-Bench (1 / 12, score 64.90), AutomationBench (1 / 12, score 50.80), DeepSWE (4 / 32, score 69.30). This page also consolidates core specs, context limits, and API pricing so you can evaluate the model from benchmark results and deployment constraints together.
Qwen3.8-Max-0902 is Alibaba Qwen's upgraded Qwen3.8-Max snapshot released on September 2, 2026. QwenCloud also lists the alias qwen3.8-max-2026-09-02. This is not a silent replacement for the earlier entry: Qwen reports separate results for 0902 and the original model, while QwenCloud exposes a dedicated model page and API identifier. DataLearner therefore preserves both releases and their historical evaluations.
The snapshot retains Qwen3.8-Max's 2.4-trillion-parameter architecture and approximately 1 million tokens of context. QwenCloud lists about 991K maximum input in non-thinking mode, about 983K maximum input in thinking mode, 131K maximum output, and a 262K maximum reasoning budget. It accepts text, image, and video input and produces text. Documented platform features include prefix completion, function calling, context caching, structured outputs, batch processing, web search, a code interpreter, and image-search tools.
Qwen says the 0902 snapshot received further post-training on Coding and Cowork data, targeting complex enterprise tasks, scientific research, and long-horizon workflows. QwenCloud describes stronger engineering-scale coding and autonomous development, steadier multi-tool orchestration and end-to-end delivery, plus refinements to chart reasoning, document parsing, and multimodal perception.
Qwen's comparison table reports gains over the original Qwen3.8-Max on Terminal-Bench 3.0 (29.0 versus 11.3), DeepSWE 1.1 (69.3 versus 56.6), NL2Repo-Bench (64.9 versus 55.9), ProgramBench (28.0 versus 10.5), SWE-Marathon (44.8 versus 39.1), MLS-Bench-Lite (50.1 versus 41.0), QwenSWEBench V2 (70.0 versus 55.1), CoWorkBench (76.1 versus 74.8), JobBench (64.0 versus 53.4), and Toolathlon Verified (73.3 versus 72.5). Reported multimodal scores include 82.7 on MMMU-Pro, 78.3 on ERQA, 80.2 Pass@3 on ClawEval-MM, and 93.8 on BabyVision with a code interpreter. These are vendor-reported results produced with different agent harnesses and tool configurations, so they should not be treated as directly comparable measurements outside the stated conditions.
Qwen3.8-Max-0902 is live on the QwenCloud API under qwen3.8-max-0902. Official prices per million tokens are $2 input and $6 output; implicit cache hits cost $0.25, explicit cache creation costs $2.50, and explicit cache reads cost $0.17.
No. The 0902 snapshot has a dedicated API identifier, model page, and evaluation column. It inherits the Qwen3.8-Max architecture but is preserved as a separate release so historical results remain intact.
QwenCloud lists a 1M-token context window, about 991K maximum input without thinking, about 983K with thinking, a 131K maximum output, and a 262K maximum reasoning budget.
Qwen reports additional post-training on Coding and Cowork data, improving engineering-scale coding, long-horizon autonomous development, enterprise and research tasks, multi-tool orchestration, and native visual understanding.
Per million tokens, standard input costs $2 and output costs $6. Implicit cache hits cost $0.25, explicit cache creation costs $2.50, and explicit cache reads cost $0.17.
Qwen currently exposes 0902 as a QwenCloud API snapshot and has not published a separate dated weight repository. The earlier Qwen3.8-2.4T-A95B repository should not be presented as a 0902 checkpoint.
Follow DataLearner on WeChat for AI model updates and research notes.
