Qwen3.8-27B Passes 7 Million Downloads Under Apache 2.0, With Training Data Undisclosed
A.I. / news
Qwen3.8-27B Passes 7 Million Downloads Under Apache 2.0, With Training Data Undisclosed
Alibaba's dense 27B model has been out since August 14. The benchmark margins over Claude Opus 4.6 are Alibaba's own, and the card says nothing about training data.
Alibaba's Qwen3.8-27B has 7,038,259 downloads and 16,631 likes on Hugging Face as of Wednesday, the highest of the 14 open-weight models on the platform's trending list that day. The dense 27-billion-parameter model was released on August 14 under the Apache 2.0 licence, which permits commercial use.
The figures come from Hugging Face's public listing and the model card. Release details are corroborated by DataNorth AI.
What the model card specifies
The model is dense, not a mixture of experts, so all 27 billion parameters are active on every token. DataNorth puts the total at 27.78 billion once the vision encoder is included. It has 64 layers and a hidden dimension of 5,120, alternating three Gated DeltaNet blocks with one Gated Attention block.
The native context window is 262,144 tokens, extendable to 1,000,000 with YaRN scaling. The model accepts text, images and video. Thinking mode is on by default, and a reasoning_effort setting takes xhigh, medium or low.
The card recommends SGLang, vLLM or TokenSpeed for serving. DataNorth says the model fits on a single high-end consumer or workstation GPU, without giving a VRAM figure, and OpenRouter lists it at $0.45 per million input tokens and $3.20 per million output tokens.
Benchmarks are Alibaba's, not independent
Every score below is reported by Alibaba on the model card. No third party is named as having reproduced them. The comparison model is Claude Opus 4.6 Max, an older release than Opus 5.5.
- Claude Opus 4.6 Max78.2 %
- Qwen3.8-27B73 %
- Muse Glimmer-30B51.7 %
Source: Qwen3.8-27B model card and DataNorth AI, accessed 2026-09-30
| Benchmark | Qwen3.8-27B | Claude Opus 4.6 Max |
|---|---|---|
| SWE-bench Pro | 61.7 | 53.4 |
| OSWorld-Verified | 84.3 | 72.7 |
| GPQA Diamond | 89.2 | 91.3 |
| Terminal Bench 2.1 | 73.0 | 78.2 |
The split is plain in the table. Qwen3.8-27B leads on the coding-agent and computer-use rows and trails on graduate-level science questions and terminal tasks. DataNorth also reports 42.2% on DeepSWE 1.1, against 13.3% for the previous version.
What is not disclosed
The card does not describe training data, and DataNorth's material does not either. A reader deciding whether the Apache 2.0 licence is enough for a regulated product has no provenance statement to check.
Alibaba has not published hosted pricing for the model, according to DataNorth. Whether 7 million downloads reflects deployments or automated pulls of quantised copies is not something the listing shows.
For other open-weight releases reviewed on this site, see the Zdtaichu 5.0 9B licence and benchmark report and the Contrastive-LM CLM 8B coverage. The next check is independent reproduction of the SWE-bench Pro and OSWorld-Verified scores.
Sources
More in A.I.
- 01GPT-Synopsys: OpenAI Gets Paid Only When the Chips It Helps Design Beat the Customer's BaselineThe Sept. 30 deal pairs an OpenAI model with Synopsys' design software, with no price, no release date and no named customer.
- 02Ataraxos Beats Stratego's Top Player 15-1-4 After Training on 16 H100s for a WeekA Nature paper from MIT, Carnegie Mellon, NYU and Stanford puts the compute bill under $8,000, against an estimated $3 million to $4.5 million for DeepMind's DeepNash.
- 03Nine Mathematicians Advising OpenAI Ask AI Labs to Stop Testing Hard Problems on Models Nobody Else Can UseThe Advisory Group on Mathematics and AI published its rules on Sept. 29: release fast, fund human understanding, disclose prompts and costs.
- 04Qwen3.8-27B Ships Under Apache 2.0 and Fits in 17GB, but Spends 160 Million Tokens Where the Median Spends 43 MillionAlibaba's open-weight model scores 52 on Artificial Analysis's Intelligence Index. Its own benchmark figures are vendor-supplied, and users report slow runs.