Qwen3.8-27B Reaches 7.1 Million Downloads Under Apache 2.0, on Benchmarks Only Alibaba Has Run
A.I. / news
Qwen3.8-27B Reaches 7.1 Million Downloads Under Apache 2.0, on Benchmarks Only Alibaba Has Run
The August 14 model beats Opus 4.6 Max on agentic coding by Alibaba's scores and trails it on reasoning. The card gives no training data and no GPU requirement.
Alibaba's Qwen3.8-27B, an Apache 2.0 open-weight model released August 14, had 7,119,569 downloads and 17,660 likes on Hugging Face as of Wednesday, and every benchmark on its card was run by Alibaba.
The model card describes a dense model with 27 billion parameters in the language model and 28 billion counting the vision encoder. It takes text, images and video and outputs text. DataNorth, which reviewed the release on August 17, puts the exact count at 27,781,427,952 parameters.
What the Qwen3.8-27B card specifies
The context window is 262,144 tokens natively, extensible to 1,000,000 with YaRN scaling. DataNorth reports that the card warns common open-source frameworks apply YaRN statically, which can reduce accuracy on shorter inputs.
The architecture repeats three Gated DeltaNet blocks for every one Gated Attention block across 64 layers. Thinking is on by default and can be switched off per request, and a reasoning_effort setting takes the values xhigh, medium and low.
The card names pre-training and post-training as stages but does not describe the data. It gives no GPU or memory requirement. It lists SGLang, vLLM, TokenSpeed and Transformers as serving options. DataNorth says Alibaba claims the model runs on a single high-end consumer or workstation GPU, without a memory figure to back it.
The licence is Apache 2.0, which allows commercial use and carries no revenue clause. A sibling release is different. According to eWeek, the 2.4-trillion-parameter Qwen3.8-2.4T-A95B, with 95 billion active, ships under a custom Max licence that requires a separate agreement from model-as-a-service businesses with more than $50 million in revenue over any 12 consecutive months.
Qwen3.8-27B benchmarks and who ran them
Alibaba scored the model against Claude Opus 4.6 Max, using the Claude Code harness at temperature 1.0, top_p 0.95 and a 256K context, according to DataNorth's reading of the card. The card also says some scores come from in-house benchmarks.
| Benchmark (Alibaba-run) | Qwen3.8-27B | Opus 4.6 Max |
|---|---|---|
| SWE-bench Pro | 61.7 | 53.4 |
| OSWorld-Verified | 84.3 | 72.7 |
| Terminal Bench 2.1 | 73.0 | 78.2 |
| GPQA Diamond | 89.2 | 91.3 |
| Humanity's Last Exam | 30.8 | 40.0 |
The model leads on agentic coding and computer use and trails on knowledge reasoning. QwenSWEBench (79.0) and CoWorkBench (70.7) are Alibaba's own tests, and DataNorth says neither has a published task list or sample size.
DataNorth also reports that Alibaba removed problematic tasks from SWE-bench Pro and re-scored all baselines on the amended set. That makes the 61.7 hard to compare with the 53.5 on the Qwen3.6-27B card, which was released in April. On DeepSWE 1.1, the 3.8 card lists 42.2 against 13.3 for Qwen3.6-27B.
No independent evaluation appears in DataNorth's write-up or on the card.
Qwen3.8-27B price, Muse Glimmer and downloads
DataNorth found no first-party Alibaba API price on August 17, when the Qwen Cloud page returned a 404. OpenRouter listed the model at $0.45 per million input tokens and $3.20 per million output tokens. Gemini 4 Argon, by comparison, launched at $2 and $10.
DataNorth says Meta released Muse Glimmer-30B under Apache 2.0 on August 10, four days earlier. On the four benchmarks where both are reported, the 27B leads, by Alibaba's figures for both. Open-weight embedding models are also moving fast, as EmbeddingGemma 2 showed.
The 7.1 million downloads count file fetches, not deployments, so they say how often the weights were pulled and nothing about who runs them. Alibaba also publishes an FP8 variant, Qwen/Qwen3.8-27B-FP8, which DataNorth says performs nearly identically to the original.
Sources
More in A.I.
- 01Torvalds Posts a Kernel Patch Series Written With Claude Code, and Sources Disagree on Its Size and SpeedupLWN counts ten patches and an order-of-magnitude gain, while It's FOSS counts eight patches and a threefold gain on one laptop.
- 02Parsimony Ranks Coding Agents by How Much Code Their SWE-bench Fixes Add, and Its Own Page Warns Smaller Is Not BetterThe MIT-licensed benchmark counts coding units in published patches without rerunning a single test, and reports that code volume tracks solve rate.
- 03OpenBMB's MiniCPM-V 4.7 Went Up With No Model Card on Oct. 6, and Now Reads Apache-2.0 With No Benchmark NumbersThe 35-billion-parameter mixture-of-experts vision model is 27 times larger than its predecessor, and its card claims a win over Qwen3.6 35B-A3B without a score.
- 04Design Creativity Bench Scores GPT-6 Astra 65.5 Against Humans' 92.3, and the Paper Never Names Its JudgeA post using Simon Willison's pelican SVGs to argue that models converge on one image cites a leaderboard whose method the arXiv page does not describe.