Qwen3.8-27B Ships Under Apache 2.0 With 262,144-Token Context, While Its 2.4T Sibling Carries a $50 Million Revenue Clause
A.I. / news
Qwen3.8-27B Ships Under Apache 2.0 With 262,144-Token Context, While Its 2.4T Sibling Carries a $50 Million Revenue Clause
The model card lists 25 benchmark rows, all Alibaba's own, and says nothing about training data or the hardware needed.
Alibaba's Qwen team released Qwen3.8-27B on Hugging Face and ModelScope in August under the Apache 2.0 licence, and the repository has since passed 18,000 likes and 7.7 million downloads. The weights download without a gate.
The model is a dense 27B-parameter vision-language model, listed as 28B parameters in BF16 on the Hub, with a native context window of 262,144 tokens that YaRN scaling extends to 1,000,000. The model card warns that static YaRN scaling may affect performance on shorter text.
What the model card says
The card describes a 64-layer hybrid of Gated DeltaNet and Gated Attention blocks, each followed by a feed-forward layer with an intermediate dimension of 17,408. It builds on the Qwen3.5 architecture. Thinking mode is on by default and can be turned off per request, according to eWeek.
Alibaba says the model runs on consumer hardware and, after quantization, on a laptop. The card lists no memory requirement. Its only deployment example is an SGLang Docker command with --gpus all and --shm-size 32g.
Training data is not described. The card names only pre-training and post-training as stages.
Benchmarks, all from Alibaba
The card compares Qwen3.8-27B with Qwen3.6-27B, Qwen3.7-Plus, Muse Glimmer-30B and Claude Opus 4.6 Max. It outscores Qwen3.6-27B on every row listed. Opus 4.6 Max leads on some text tests.
| Benchmark | Qwen3.8-27B | Qwen3.6-27B |
|---|---|---|
| Terminal Bench 2.1 | 73.0 | 63.4 |
| SWE-bench Pro | 61.7 | 53.5 |
| OSWorld-Verified | 84.3 | 63.9 |
On Terminal Bench 2.1, Opus 4.6 Max scores 78.2 against Qwen's 73.0, and on GPQA Diamond it scores 91.3 against 89.2. MLQ.ai, quoted by eWeek, said that the release lacks a complete like-for-like comparison with earlier Qwen models and rival open models.
- Claude Opus 4.6 Max78.2 points
- Qwen3.8-27B73 points
- Qwen3.6-27B63.4 points
Source: Qwen3.8-27B model card on Hugging Face, accessed 2026-10-26 (vendor-supplied)
The 2.4T release has a different licence
Alibaba released Qwen3.8-2.4T-A95B on the same day, a mixture-of-experts model with 2.4 trillion total and 95 billion active parameters. According to eWeek it carries a custom Qwen3.8-Max licence.
Under it, a business running a qualifying model-as-a-service or AI work assistant needs a separate licence once aggregate revenue exceeds $50 million in any 12 consecutive months. Internal use is not covered by the requirement. The 27B model has no such clause.
Alibaba's cloud API lists $2 per million input tokens, per eWeek. The Terminal has not seen an independent evaluation of either model.
The 27B model's popularity sits against a market where DeepSeek's V4.1-Flash-1017 is priced far lower on hosted APIs, and where Cerebras serves Kimi K3 as a rival large open model. Alibaba has not said when a successor to the 27B size will follow.
Sources
More in A.I.
- 01OpenAI Cut Luna 80% in One Step. Anthropic Has Cut 20% at a Time.Since June 26 both labs have repriced their model lines, and the dated record shows two different methods, with DeepSeek still far below either.
- 02Microsoft's Decision-1 Is a Post-Trained Qwen3.5-9B at $0.042 Per Million Tokens, With Self-Reported BenchmarksThe model scores fixed answer options instead of writing text, and the launch post does not say whether the weights can be downloaded.
- 03Gemini 4 Argon Reached General Availability October 13 at a $2 Promotional Price, After a Cyber-Partners-First LaunchGoogle's own post gives four benchmark scores, concedes two coding tests to rivals in Reuters' account, and sets a $4 input price from November 30.
- 04Eka's 1,500 Picks an Hour Is a Lab Number, and Its Own Investors Say SoThe Cambridge, Mass. startup raised $150 million for a vision-force-action robot model. Its headline throughput comes from induction and bin-picking tests, and customer pilots are unnamed.