Qwen3.8-27B's Model Card: Apache 2.0, a 1M-Token Ceiling and Scores Run Through Claude Code
A.I. / explainer
Qwen3.8-27B's Model Card: Apache 2.0, a 1M-Token Ceiling and Scores Run Through Claude Code
Alibaba's 27-billion-parameter dense model beats Claude Opus 4.6 Max on SWE-bench Pro by its own count, on a task set Alibaba modified.
Alibaba's Qwen team released Qwen3.8-27B on August 14, 2026, a 27-billion-parameter dense vision-language model under the Apache 2.0 licence. Its model card on Hugging Face reports 61.7 on SWE-bench Pro, against 53.4 for Anthropic's Claude Opus 4.6 Max. Every score in the card is Alibaba's own, and the card cites no outside reproduction.
What Qwen3.8-27B is, and what it needs
The checkpoint has 64 layers and a hidden dimension of 5,120. The card describes a layout of 16 repeating blocks, each with three Gated DeltaNet layers followed by one Gated Attention layer. Hugging Face lists it at about 28 billion parameters in BF16, because the vision encoder is counted. DataNorth puts the exact count at 27,781,427,952.
The native context window is 262,144 tokens. Alibaba says it extends to 1,000,000 tokens with YaRN. The card carries a warning that enabling YaRN in common open-source frameworks can reduce accuracy on shorter inputs.
The card gives no GPU or memory figure. It names SGLang, vLLM and TokenSpeed as serving frameworks and ships Docker examples. An FP8 variant, Qwen3.8-27B-FP8, is published separately. DataNorth says it fits on one high-end consumer or workstation GPU but gives no VRAM number either.
The licence is Apache 2.0. DataNorth confirmed it against the licence file and reads it as allowing commercial use with no revenue share.
The scores, and how they were produced
The card compares the model with Qwen3.6-27B, Qwen3.7-Plus, Meta's Muse Glimmer-30B and Claude Opus 4.6 Max. Alibaba ran models through the Claude Code harness at temperature 1.0, top_p 0.95 and a 256K context window. The exception is Opus 4.6 Max on SWE-bench Pro, where Alibaba used Anthropic's officially reported score.
Alibaba also changed the SWE-bench Pro task set and corrected some MathVision and CharXiv annotations, according to DataNorth. Those scores are not directly comparable to the unmodified benchmarks.
- Qwen3.8-27B61.7 %
- Qwen3.7-Plus57.6 %
- Claude Opus 4.6 Max53.4 %
- Muse Glimmer-30B51.2 %
Source: Qwen/Qwen3.8-27B model card and DataNorth, accessed 2026-11-13
The card does not win every row. Opus 4.6 Max leads on three of the headline benchmarks.
| Benchmark | Qwen3.8-27B | Opus 4.6 Max |
|---|---|---|
| Terminal Bench 2.1 | 73.0 | 78.2 |
| GPQA Diamond | 89.2 | 91.3 |
| Humanity's Last Exam | 30.8 | 40.0 |
| OSWorld-Verified | 84.3 | 72.7 |
Two more rows, QwenSWEBench at 79.0 and CoWorkBench at 70.7, are Alibaba's in-house benchmarks with little public detail. DataNorth said they cannot be independently checked. The previous Qwen3.6-27B scored 49.3 on QwenSWEBench, per the card.
Thinking controls and the serving recipe
Thinking is on by default and can be switched off per request. The card exposes a reasoning_effort setting that accepts xhigh, medium or low, with xhigh as the default, and a preserve_thinking flag that is also on by default.
Alibaba's recommended sampling differs by mode. Thinking mode uses temperature 1.0, top_p 0.95 and top_k 20. Non-thinking mode uses temperature 0.7, top_p 0.8, top_k 20 and a presence penalty of 1.5. The card says the model is trained with multi-token prediction over several steps, which serving stacks can use for speculative decoding.
Against Meta's Muse Glimmer-30B, released August 10, 2026, DataNorth reports Alibaba leading on all four shared benchmarks, including Terminal Bench 2.1 at 73.0 to 51.7.
Where it already shows up in other releases
The weights are being reused. Nous Research's Arete-27B is a fine-tune of Qwen3.8-27B, and the Edge0-145B model card names a Qwen model as its base. Both lean on the Apache 2.0 terms.
The card says nothing about training data or compute. Its training section is the single line "Pre-training & Post-training." The citation block is titled for Qwen3.8-Max, a 2.4-trillion-parameter mixture-of-experts model whose open weights are text-only, which does not match the name of this model.
Alibaba had not published official API pricing as of DataNorth's August 17 report, when its Qwen Cloud page returned a 404. OpenRouter listed the model at $0.45 per million input tokens and $3.20 per million output tokens. No independent run of the benchmarks had appeared as of November 13.
Sources
More in A.I.
- 01Apple's Harbor App Would Cluster Thunderbolt 5 Macs for Local AI, per a LeakerThe only source for the unreleased app is leaked text attributed to Aaron Perris; Apple has announced nothing, though the clustering layer underneath it is documented.
- 02Nous Research's Arete-27B Cuts Propaganda Answers From 22% to 0.9% on Its Own TestThe Qwen3.8-27B fine-tune is Apache-2.0 and open on Hugging Face, but the evaluation behind the headline number was built by its trainers.
- 03Edge0-145B's Model Card Names Qwen as Its Base, and the Launch Post Never DoesThe 145-billion-parameter open-weights release says it runs on a laptop, but its licence is two documents and its benchmarks are all Edge0's own.
- 04Chrome's Modern Web Guidance Reaches 9,200 Stars While Its Docs and README Count 100 and 203 Use CasesThe agent skill set from the Chrome team reports a 37-point average gain in best-practice adherence, but publishes the method only on a dashboard.