Qwen3.8-Max (open weights: Qwen3.8-2.4T-A95B)

2026-08-13 · Publisher : Alibaba · Domain : Artificial Intelligence

Alibaba launched Qwen3.8-Max through its API on August 3, 2026, then released its weights (Qwen3.8-2.4T-A95B, 95 billion active parameters) under a house license, a first for a Max-class model.

### Context Qwen's previous Max and Plus models (Qwen3.7-Max and Qwen3.7-Plus, released in May and June 2026 according to the press) were available only through an API. Qwen3.8 breaks with that practice. This entry is here because the open/closed gap can be read within a single family: the hosted Qwen3.8-Max remains a proprietary service with more features, while the weights under a restrictive license can be downloaded. ### What the model brings - Open Max class: the model card states that Qwen3.8 brings, for the first time, a Max-class model to an open release. - Targets: coding, professional work, research and long-horizon agentic tasks, with better end-to-end execution (vendor announcement). - Thinking control: `reasoning_effort` sets the depth, `preserve_thinking` keeps the reasoning of previous messages. - A Terminal Bench 2.1 score of 86.6, two points below the best of the three closed competitors in the model card's table. - The official Qwen3.8-Max version adds vision input, a no-thinking mode, 1M of context by default and built-in tools, all absent from the public weights. ### Architecture and specifications - 2.4T parameters in total, 95B active; 92 layers; layout 23 x (3 x (Gated DeltaNet then MoE) then 1 x (Gated Attention then MoE)). - MoE: 512 experts, 10 routed plus 1 shared; hidden dimension 8192; multi-token prediction (MTP) training. - Context: 262,144 tokens natively, extensible up to 1,010,000. - License: Qwen3.8-Max License, MIT-style with conditions. Commercial use is free, with two conditions: display the model name if the product exceeds 100 million monthly users or USD 20 M of monthly revenue, and obtain a separate license from Qwen to operate the model as an inference service (Model as a Service) or an AI work assistant (code, office productivity) if aggregate revenue exceeds USD 50 M over 12 months. Internal use is exempt. - Weights: Hugging Face, Qwen/Qwen3.8-2.4T-A95B (and FP8 version). Knowledge cutoff: not disclosed. ### Cost and access | Item | Value | |---|---| | Qwen3.8-Max API, input (USD / M tokens) | 2 | | Qwen3.8-Max API, output (USD / M tokens) | 6 | | Implicit cache / explicit read | 0.25 / 0.17 | | API context | 1M tokens (991K input, 131K output) | | Self-hosting | Not disclosed; a 2.4T model requires a multi-node deployment (FP8 weights reduce memory) | ### Benchmarks | Test | Opus 4.8 | Fable 5 | GPT 5.6 Sol (max) | Qwen3.7-Max (closed) | Qwen3.8-Max | |---|---|---|---|---|---| | Terminal Bench 2.1 | 84.6 | 84.6 | 88.8 | 74.5 | 86.6 | | SWE-bench Pro | 69.2 | 80.0 | 64.6 | 60.6 | 67.7 | | DeepSWE 1.1 | 59.0 | 70.0 | 73.0 | 21.6 | 56.6 | | GPQA Diamond | 92.0 | 92.6 | 94.1 | 92.4 | 92.6 | | HLE (no tools) | 45.7 | 53.3 | 47.2 | 41.4 | 43.6 | | HLE with tools | 57.9 | 64.5 | 58.0 | 53.5 | 56.2 | | IFBench | 62.2 | 63.5 | 72.7 | 79.1 | 82.8 | Source: Hugging Face model card for Qwen3.8-2.4T-A95B, scores published by the vendor (Qwen3.8-Max is the hosted version of the same model). ### Versus closed models On this table, Qwen3.8-Max is ahead of Opus 4.8 on Terminal Bench 2.1 (86.6 vs 84.6) and on IFBench (82.8, the best score in the row), but remains behind GPT 5.6 Sol (88.8) on Terminal Bench 2.1. It trails Fable 5 on SWE-bench Pro (67.7 vs 80.0), on HLE (43.6 vs 53.3) and on DeepSWE 1.1 (56.6 vs 70.0). It improves sharply on its closed predecessor Qwen3.7-Max (SWE-bench Pro 60.6 then 67.7; DeepSWE 21.6 then 56.6). The announced API price is 2 and 6 USD per million tokens; the prices of the three closed competitors are not in the source and are not compared here. For the open/closed choice, the contribution is twofold: downloadable weights (self-hosting, sovereignty) and a measured gap of a few points on coding, but under a license that requires a separate agreement for inference service providers beyond USD 50 M of revenue. ### Key takeaways A Max-class model is released with open weights, which is new for Qwen, but its license is not Apache 2.0 and commercial exploitation as a service is restricted. The hosted Qwen3.8-Max service remains proprietary.

Tags : LLM, Poids ouverts, Alibaba, Qwen, MoE, Licence Qwen3.8-Max, Version API propriétaire (Max)