August 9, 2026
Alibaba’s Qwen team published the weights of Qwen3.8-Max on Hugging Face on August 9, six days after launching the model through Alibaba Cloud Model Studio. The open checkpoint, named Qwen3.8-2.4T-A95B, takes text only and supports 262,144 tokens natively, extensible to 1,010,000, according to its model card.
Takeaway points
- Hugging Face counts 2,446,182,725,504 parameters in the open weights; the config lists 512 experts with 10 active per token across 92 layers.
- The model card calls the hosted Qwen3.8-Max the official version, with vision input, non-thinking support, 1 million tokens by default and built-in tools.
- The Max weights carry a custom Qwen3.8-Max License, while the smaller Qwen3.8-27B uses Apache 2.0, a standard open-source licence.
Qwen3.8-Max parameters, architecture and context length
“For the first time, Qwen3.8 brings a Qwen-Max-class model to open release,” the model card says. Alibaba said on August 3 that Qwen3.8-Max activates 95 billion of its 2.4 trillion parameters and ranks fifth in Text Arena, second in Vision Arena and fourth in Frontend Code Arena, and that the weights were “scheduled for release next week.” The checkpoint’s name, 2.4T-A95B, matches those 95 billion active parameters.
The card says the model is built on the architecture of Qwen3.5. Its 92 layers run in 23 repeating groups, each made of three Gated DeltaNet layers, a form of linear attention, and one Gated Attention layer, with every layer followed by a mixture-of-experts block. Each token uses 10 routed experts and 1 shared expert out of 512, with an expert intermediate size of 2,048 and a hidden size of 8,192. The model was trained for multi-token prediction, and the card says it runs on vLLM, SGLang and TokenSpeed. The Gated DeltaNet layers use 128 linear attention heads for values and 16 for queries and keys, and the Gated Attention layers use 64 query heads and 4 key-value heads, over a padded vocabulary of 248,320 tokens. Reasoning depth is set with a reasoning_effort parameter, and reasoning from earlier messages can be kept with a preserve_thinking setting.
The repository history shows the weights uploaded on August 9, the model card on August 11 and the licence on August 12. Hugging Face records 55,288 downloads of the checkpoint. A mixture-of-experts model uses only part of its weights for each token but still has to hold all of them in memory, as StrideNote measured in MoE on a Mac.
Qwen3.8-Max benchmarks against GPT-5.6 Sol and Claude
The model card compares Qwen3.8-Max with Claude Opus 4.8, Claude Fable 5, GPT-5.6 Sol at maximum settings and Alibaba’s previous flagship, Qwen3.7-Max. On Terminal Bench 2.1 it scores 86.6, against 88.8 for GPT-5.6 Sol and 84.6 for both Anthropic models. On SWE-bench Pro it scores 67.7, against 80.0 for Fable 5. On DeepSWE 1.1 it scores 56.6, against 73.0 for GPT-5.6 Sol, but up from 21.6 for Qwen3.7-Max.
It leads on some tests. On PaperBench it scores 93.0, against 90.5 for GPT-5.6 Sol. On IFBench, an instruction-following test, it scores 82.8 against 72.7, and on HealthBench 60.2 against 55.3. On GPQA Diamond it matches Fable 5 at 92.6.
The footnotes qualify these numbers. Alibaba says the Fable 5 results “may involve fallbacks,” ran Terminal Bench with the Claude Code harness under a 5-hour timeout, and corrected problematic SWE-bench Pro tasks before evaluating every model on the refined set. Several rows, such as QwenSWEBench, are Alibaba’s own in-house benchmarks.
Qwen3.8-Max license terms for commercial use
The licence opens with a broad grant in the style of the MIT licence, covering use, modification, hosting, fine-tuning and sale. Two conditions follow. Any commercial product built on the model with more than 100 million monthly active users, or more than $20 million in monthly revenue, must display the model’s name prominently. A business that sells model inference as a service, or an “AI Work Assistant,” with more than $50 million in revenue over any 12 months must obtain a separate licence from Qwen before commercial use.
The licence defines an AI Work Assistant as an independent product mainly for AI-assisted coding or office work, citing Alibaba’s own Qoder and QwenWork as examples, and excludes single-purpose tools such as translators. Internal use that does not expose the model or its outputs to third parties is exempt. The revenue threshold is higher than the $20 million in Moonshot’s Kimi K3 License, released in July.
Qwen3.8-27B and the smaller Qwen3.8 models
Qwen3.8-27B, first committed on August 5, is a dense model with 27,781,427,952 parameters by Hugging Face’s count, under Apache 2.0. Its card describes a vision-language model that understands images and hour-scale videos, supports 262,144 tokens natively and up to 1,000,000, and keeps thinking on by default with an option to disable it per request. The card lists 64 layers and a hidden size of 5,120, and compares the model with Qwen3.7-Plus, Muse Glimmer-30B and Claude Opus 4.6 Max as well as its predecessor. Against that predecessor, Qwen3.6-27B, the card reports 73.0 against 63.4 on Terminal Bench 2.1 and 42.2 against 13.3 on DeepSWE 1.1. It has been downloaded 7,322,476 times, and its size class is covered in StrideNote’s local LLM RAM sizing guide.
A third model, Qwen3.8-Flash-Next, was uploaded on August 26 with about 180 billion parameters under a separate Qwen Community License 1.0. Its card calls it an “experimental preview of the architecture that will underpin Qwen4,” and its config names the architecture Qwen4ExpForConditionalGeneration. The card lists a new sparse attention scheme, Qwen Sparse Attention, which works on small blocks of tokens rather than individual ones, along with a gated residual design and n-gram embeddings. Ollama added Qwen3.8 Flash Next to its MLX engine for Macs in version 0.33.1 on the same day, according to its release notes.
What Alibaba has not said about Qwen3.8-Max
Alibaba’s announcement does not give API prices, and the model card does not say whether the hosted version’s vision input will reach the open weights. The Flash-Next card calls Qwen4 the next architecture but gives no date for it.
Sources: Alibaba Group; Qwen3.8-2.4T-A95B model card; Qwen3.8-Max License; its config; Qwen3.8-27B model card; Qwen3.8-Flash-Next model card; Ollama release notes.