Qwen3.8-27B Has 6.9 Million Downloads on Hugging Face. Artificial Analysis Rates It Slow and Verbose
A.I. / news
Qwen3.8-27B Has 6.9 Million Downloads on Hugging Face. Artificial Analysis Rates It Slow and Verbose
Alibaba's Apache 2.0 dense model beats its size class on an independent index of 20, and takes 38.8 minutes a task against a 12.6-minute median.

Alibaba's Qwen3.8-27B, an open-weight model released on Aug. 14, shows 6,942,530 downloads and 20,405 likes on its Hugging Face page as of Tuesday. Independent tester Artificial Analysis rates it well above similar-size open models on intelligence and well below them on speed.
The weights are ungated and carry an Apache 2.0 licence, which permits commercial use. eWeek reported the release on Aug. 17.
What the model card says
Qwen3.8-27B is a dense model with 27B parameters, 64 layers and a native vision encoder for images and video. Its layout alternates three Gated DeltaNet blocks with one Gated Attention block. Context is 262,144 tokens natively and extensible to 1,000,000.
Thinking is on by default at a reasoning_effort of xhigh, with medium and low available. The card carries its own warning: in multi-turn agentic tasks, lower reasoning effort does not always shorten total completion time, because it can cause more failures.
Vendor numbers
The card's benchmark table compares the model with Qwen3.6-27B, Qwen3.7-Plus, Muse Glimmer-30B and Opus4.6 Max. These are Alibaba's figures. Several come from in-house benchmarks, QwenSWEBench and CoWorkBench among them.
| Benchmark (Alibaba-run) | Qwen3.8-27B | Opus4.6 Max |
|---|---|---|
| Terminal Bench 2.1 | 73.0 | 78.2 |
| SWE-bench Pro | 61.7 | 53.4 |
| OSWorld-Verified | 84.3 | 72.7 |
| GPQA Diamond | 89.2 | 91.3 |
The card says the Opus4.6 Max SWE-bench Pro score is the officially reported one, while the other models ran on the Claude Code harness. eWeek quoted MLQ.ai saying the release materials "do not provide a complete, like-for-like benchmark comparison."
What Artificial Analysis measured
Artificial Analysis scores the model 20 on its Intelligence Index at xhigh, 17 at medium, 16 at low and 10 with reasoning off. The median for similar-size open models is 6.
- Qwen3.8-27B, xhigh38.8 min
- Qwen3.8-27B, medium30.7 min
- Qwen3.8-27B, low28.2 min
- Qwen3.8-27B, non-reasoning18.9 min
- Median, similar-size open models12.6 min
Source: Artificial Analysis, Qwen3.8 27B page, accessed 2026-11-17
Time to first answer token at xhigh is 124.37 seconds against a 4.99-second median. The model produced 128M tokens on the index against a 77.5M median. Artificial Analysis ranks it eighth of nine on cost per task and ninth of nine on speed, at 47 output tokens per second. Cost per task is $3.53.
Its listed Alibaba API price is $0.50 per million input tokens and $3.00 per million output tokens, against a $0.13 and $0.30 median.
Where this fits among open releases
The pattern of strong scores and heavy token use is the cost an agent team pays for reasoning at the default setting. Thinking Machines Lab's Inkling 2 Small is another Apache 2.0 release, and benchmark claims of this kind are the subject of Nvidia's SkillEvaluator report.
Alibaba's larger Qwen3.8-2.4T-A95B has 2.4 trillion total and 95 billion active parameters and uses a separate Qwen3.8-Max licence. eWeek reported that it requires a separate licence for qualifying model-as-a-service businesses above $50 million in revenue over 12 months. The 27B model carries no such condition.
Alibaba said on the card that a hosted version with 1M context by default is coming to Qwen Cloud. The card gives no date.
Sources
More in A.I.
- 01Runta's 27,280-Run Harness Eval Finds the Model Explains 92.8% of Agent CostFrontierHarness v1.1 ran 22 coding-agent harnesses against five open models, and Codex CLI went from first place to sixteenth when the model changed.
- 02Claude Code Team Publishes Its Verification Rules: 39 Checks, a Three-Round Limit and Observe ModeA Nov. 16 post says about 90% of the team's code is written and reviewed by Claude, and lays out the gates that keep it shippable.
- 03Inkling 2 Small Weights Land on Hugging Face Under Apache 2.0, and the Card Asks for 500GB of VRAMThinking Machines opened the weights a week after launching on Tinker. The safety section still says release is pending.
- 04Fed Governor Barr Says AI Agents Could Accelerate Bank Runs That Unfold in MinutesIn a Nov. 16 speech in Milan, Barr also put AI firms' off-balance-sheet liabilities above $3 trillion and proposed no new rules.