Qwen3.8-27B Hits 6.3 Million Downloads on Alibaba's Own Benchmark Claims
A.I. / news
Qwen3.8-27B Hits 6.3 Million Downloads on Alibaba's Own Benchmark Claims
The weights are Apache 2.0 and ungated, but the card says nothing about training data, languages or the GPU needed to run them.

Qwen3.8-27B, Alibaba's open-weight vision-language model, listed 20,878 likes and 6,297,396 downloads on Hugging Face on Sunday, Nov. 22, according to the site's API record. The repository was created Aug. 5 and last modified Nov. 7.
The weights are ungated and carry an Apache-2.0 licence, which permits commercial use. The model card is the only primary source, and every benchmark on it is Alibaba's own.
What the card says about the model
Qwen3.8-27B is a dense model, meaning all parameters run on every token. The card says 27B; the Safetensors metadata totals about 27.8 billion in BF16. It stacks 64 layers in a hybrid pattern: three Gated DeltaNet layers, then one Gated Attention layer, repeated 16 times.
The native context window is 262,144 tokens. YaRN scaling stretches it to 1,000,000, but the card warns the scaling is static and can hurt short-text performance. Thinking mode is on by default, and reasoning_effort accepts xhigh (the default), medium or low.
The card does not give a release date, a training-data description, a language list or a GPU memory figure. It names vLLM, SGLang and TokenSpeed as serving options and shows a Docker example with --shm-size 32g.
The benchmark table is a vendor table
The card compares the model with Qwen3.6-27B, its predecessor. Alibaba's figures are below. The Terminal did not run any of them.
| Benchmark (vendor-supplied) | Qwen3.8-27B | Qwen3.6-27B |
|---|---|---|
| Terminal Bench 2.1 | 73.0 | 63.4 |
| SWE-bench Pro | 61.7 | 53.5 |
| QwenSWEBench | 79.0 | 49.3 |
| OSWorld-Verified | 84.3 | 63.9 |
| GPQA Diamond | 89.2 | 87.8 |
The largest jump, 29.7 points, is on QwenSWEBench, a benchmark carrying the developer's own name. The other rows move less: 9.6 points on Terminal Bench 2.1 and 1.4 on GPQA Diamond.
- Claude Opus 4.6 Max78.2 points
- Qwen3.8-27B73 points
- Qwen3.6-27B63.4 points
Source: Qwen3.8-27B model card (vendor-supplied), accessed 2026-11-22
The same table lists Claude Opus 4.6 Max ahead on both rows where it appears: 78.2 against 73.0 on Terminal Bench, and 91.3 against 89.2 on GPQA Diamond. The card's wording is that Qwen3.8-27B is "the most capable generation in the Qwen open-model family to date," which the table supports only against Qwen's own earlier models.
The larger sibling has a licence limit
Alibaba released a second model, Qwen3.8-2.4T-A95B, with 2.4 trillion total and 95 billion active parameters. eWeek reported on Aug. 17 that it sits under a custom Qwen3.8-Max licence. Businesses running a qualifying model-as-a-service or AI work assistant need a separate licence above $50 million in revenue over any 12 consecutive months. The 27B model has no such clause.
eWeek, which put the 27B release on "Friday" in its Aug. 17 report, also relayed a caveat from MLQ.ai: the release materials lack a complete, like-for-like benchmark comparison.
Derivatives are already appearing

A repository called harness-hub/Qwen3.8-27B-pi-adapted, created Nov. 16, has 985 likes and 2,831 downloads. Its listing says nothing that this article could verify about what the adaptation changes, so it counts as a sign of interest and not of quality.
Stars and downloads measure interest. Hugging Face's counter does not separate repeat pulls from CI jobs, and Alibaba says its Qwen family has passed 3 billion downloads across more than 460 models.
Another Apache 2.0 release from the same weeks is covered in Thinking Machines Ships Inkling 2 Mini. Terminal Bench scores matter most to the coding-agent tools such as the one in OpenAI Says Codex Cannot Run in Docker for iOS Builds.
The card sets different sampling for the two modes: temperature 1.0, top_p 0.95 for thinking, and temperature 0.7, top_p 0.80, presence_penalty 1.5 for instruct.
Sources
More in A.I.
- 01Anthropic's Levent Alpöge Claims a Claude-Assisted Riemann Proof, and Lean Checks Are InImperial's Kevin Buzzard and MIT's Andrew Sutherland have examined the formalization, according to Scientific American. No journal review has been reported.
- 02Thinking Machines Ships Inkling 2 Mini, a 26.4B Apache 2.0 Model Sized for One DGX SparkThe model card lists 3.4B active parameters, a 1M-token window and a 20GB checkpoint, but does not say who ran most of its benchmark rows.
- 03Anthropic Closed Its Debut at $97.17, 17% Over the $83 IPO Price, as Nasdaq Fell 2%The $81.34 billion offering opened 34% higher, then gave back most of the gain while Nvidia, AMD and Micron dropped.
- 04Google's Gemini Tool Retrieval Preview Adds 26 Tokens Per Idle Tool and Publishes No Savings FigureAnthropic and OpenAI each reported token reductions for the same defer_loading flag. Google's November 19 post reports none.