Qwen3.8-27B Has 8 Million Downloads and Only Alibaba's Own Benchmarks
A.I. / news
Qwen3.8-27B Has 8 Million Downloads and Only Alibaba's Own Benchmarks
The Apache 2.0 model's SWE-bench Pro lead over Claude Opus 4.6 Max rests on a benchmark Alibaba corrected itself, and no outside party has scored it.
Alibaba's Qwen3.8-27B, an Apache 2.0 open-weight model released on Aug. 14, 2026, recorded 8,014,877 Hugging Face downloads in the last month and has 19,000 likes. Every benchmark on its model card was run by Alibaba.
That includes a SWE-bench Pro score of 61.7 against 53.4 for Claude Opus 4.6 Max, a result built on a version of the benchmark Alibaba itself corrected.
What the model card says
The Hugging Face card describes a 27-billion-parameter dense model with a vision encoder. It uses a hybrid of Gated DeltaNet and Gated Attention blocks across 64 layers. The native context window is 262,144 tokens, extendable to 1 million with YaRN scaling. The licence is Apache 2.0 and the card describes no access gate. It says nothing about training data.
The card gives no memory requirement. DataNorth counts 27,781,427,952 parameters in the BF16 checkpoint and puts the weights alone at roughly 55 GB, its own arithmetic at 2 bytes a parameter, before cache and overhead. It also warns that early coverage wrongly called these weights text-only. That was true of Qwen3.8-Max, not this model.
Vendor-supplied scores
All figures in the table are from Alibaba's card. Higher is better.
| Benchmark | Qwen3.8-27B | Opus 4.6 Max | Qwen3.7-Plus |
|---|---|---|---|
| Terminal Bench 2.1 | 73.0 | 78.2 | 64.0 |
| SWE-bench Pro | 61.7 | 53.4 | 57.6 |
| GPQA Diamond | 89.2 | 91.3 | 90.3 |
| HLE | 30.8 | 40.0 | 34.7 |
| OSWorld-Verified | 84.3 | 72.7 | 73.3 |
- Qwen3.8-27B61.7 score
- Qwen3.7-Plus57.6 score
- Qwen3.6-27B53.5 score
- Claude Opus 4.6 Max53.4 score
- Muse Glimmer-30B51.2 score
Source: Qwen3.8-27B Hugging Face model card, accessed 2026-11-03
The model trails Opus 4.6 Max on Terminal Bench 2.1, GPQA Diamond and HLE. It leads on the screen-driven agent tasks, with 84.3 on OSWorld-Verified against 72.7.
What weakens the comparison
DataNorth, which says nobody independent has scored the model, lists the caveats. Alibaba corrected tasks in SWE-bench Pro and re-ran baselines on the amended set, so the scores are not comparable with the public benchmark. It also fixed annotations in MathVision and CharXiv.
Two headline rows, QwenSWEBench at 79.0 and CoWorkBench at 70.7, are in-house tests with no published task list or sample size. Alibaba's harness was Claude Code at temperature 1.0 and top_p 0.95 with a 256K context.
The comparison model is also older. The card measures against Opus 4.6 Max, while the site has covered Claude Opus 5.5, a newer release. No Qwen3.8-27B result against it appears on the card.
Cost and access
Alibaba has no official API price for the model. DataNorth found the Qwen Cloud page returned a 404 on Aug. 17 and saw OpenRouter listing $0.45 per million input tokens and $3.20 per million output. The card recommends SGLang, vLLM or TokenSpeed for serving. It warns that common frameworks apply YaRN statically, which can cost accuracy on short inputs.
Downloads are not adoption. The site's coverage of Heretic, an abliteration tool for open models shows what people do with open weights once they have them.
An independent run of the SWE-bench Pro and OSWorld-Verified numbers would settle most of this. None is listed on the card.
Sources
More in A.I.
- 01Anthropic's Q3 Revenue Gain Fell to $5.2 Billion From $6.9 Billion in Q2Figures shown to prospective investors on Oct. 14 put the quarter at $16.8 billion, with an 11% adjusted operating margin Anthropic has not defined.
- 02GPT-Rosalind-6 Reuses GPT-6 Astra's Weights, Cuts Biology Refusals to 9.4% and Stays Behind a GateOpenAI's Oct. 30 system card addendum says no training was added, so the difference from Astra is deployment controls and who is allowed in.
- 03Architect Says Claude Opus 5.5 Drew a House Set Good Enough for a Permit After Two Redline RoundsStephen Gallagher's informal test of 13 models gave Opus 5.5 the most human feedback, and the drawings still missed a required second exit.
- 0450 GPT Agents in Minecraft Built a Ration Ledger, and 34 Died Doing ItY Combinator's Labs group and Olam Labs ran GPT-6.1 Sol and Claude Opus 5.5 swarms through Minecraft and Mindustry. GPT won both head-to-head games, but each result comes from one harness and one setup.