What Qwen 3.8 actually shipped
Qwen’s August flagship is two products, not one. Qwen3.8-Max is the 2.4T Cloud API. The open-weight twin, Qwen3.8-2.4T-A95B, is text-only. A separate 27B vision model is the one you can actually run locally.
Hyuk · · 3 min read · AI
The Qwen team dated Qwen3.8-Max on August 2, 2026. Alibaba’s corporate note and Qwen Cloud list the production API on August 3. The headline size is 2.4 trillion parameters, 95 billion active.
That number describes two products. Qwen3.8-Max is the Cloud SKU. Qwen3.8-2.4T-A95B is the open-weight checkpoint. They share a backbone. They do not share every feature.
The two large SKUs
The Hugging Face card for Qwen/Qwen3.8-2.4T-A95B lists a Mixture-of-Experts with 512 experts and 10 routed plus 1 shared active. Native context is 262,144 tokens, extensible to 1,010,000.
Qwen Cloud’s Max page bills a 1M context window by default, with a 991K input cap and 131K output. It also accepts image, text, and video input.
The open-weight card is explicit the other way. Qwen3.8-2.4T-A95B is text-only. Thinking is required. You cannot turn it off. Multimodal input is not supported.
What Qwen’s own table says
The 2.4T card publishes Qwen’s scores against Qwen3.7-Max and several closed systems. A few of those rows:
| Benchmark | Qwen3.7-Max | Qwen3.8-Max | GPT-5.6 Sol (max) |
|---|---|---|---|
| Terminal Bench 2.1 | 74.5 | 86.6 | 88.8 |
| SWE-bench Pro | 60.6 | 67.7 | 64.6 |
| PaperBench | 64.8 | 93.0 | 90.5 |
| CoWorkBench | 64.6 | 74.8 | 71.5 |
| IFBench | 79.1 | 82.8 | 72.7 |
| GPQA Diamond | 92.4 | 92.6 | 94.1 |
| PRBench-Finance | 46.8 | 58.3 | 55.5 |
Those are Qwen’s numbers. The footnotes say the harness is not the same for every column. Terminal Bench 2.1 is Claude Code for Qwen, and “best published score” for the others. Several sets — CoWorkBench, QwenSWEBench, QwenQoderBench — are in-house.
The 27B you can actually host
Qwen/Qwen3.8-27B is a dense 27B vision-language model under Apache 2.0. Native context is 262,144, extensible to 1,000,000. The card says it reads images and video. Thinking is on by default and can be turned off. reasoning_effort is xhigh, medium, or low. preserve_thinking keeps earlier reasoning turns.
Qwen’s 27B table against the previous 27B:
| Benchmark | Qwen3.6-27B | Qwen3.8-27B |
|---|---|---|
| Terminal Bench 2.1 | 63.4 | 73.0 |
| SWE-bench Pro | 53.5 | 61.7 |
| DeepSWE 1.1 | 13.3 | 42.2 |
| QwenSWEBench | 49.3 | 79.0 |
| CoWorkBench | 61.0 | 70.7 |
| LiveCodeBench v6 | 83.9 | 90.3 |
The same card lists multimodal agent scores of 84.3 on OSWorld-Verified, 64.8 on WebArena-Verified, and 81.9 on AndroidWorld. Those are also Qwen’s evaluations.
What it costs on Qwen Cloud
Qwen Cloud and Alibaba Cloud Model Studio both list qwen3.8-max at $2 per 1M input tokens and $6 per 1M output. Implicit-cache input is $0.25. Explicit cache creation is $2.50. Explicit cache read is $0.17.
Built-in tools on that page include a code interpreter, web search, and web extract.
The license that is not Apache
The 27B weights are Apache 2.0. The 2.4T weights use the Qwen3.8-Max License. The grant is broad until two commercial gates fire: a name-display rule above 100 million monthly users or $20 million monthly revenue, and a separate Qwen license for Model-as-a-Service or AI work-assistant businesses once trailing-twelve-month revenue exceeds $50 million.
The useful split is Cloud Max versus the two Hugging Face dumps. One API speaks vision and tools. One 2.4T dump does not. The 27B dump is the small vision model.
Source cutoff: Qwen blog, Alibaba launch note, Qwen Cloud Max page, Alibaba Cloud pricing, and the Hugging Face 2.4T and 27B cards, as retrieved August 17, 2026.
Sources
Related
Comments
Comments are off until Giscus is configured.