Qwen3.8-27B Is Apache 2.0, but Artificial Analysis Clocks It at 200M Tokens
A.I. / news
Qwen3.8-27B Is Apache 2.0, but Artificial Analysis Clocks It at 200M Tokens
Alibaba's card reports 73.0 on Terminal Bench 2.1, while the independent index shows the model using 2.4 times the median output tokens to finish its test set.
Qwen3.8-27B, a 27-billion-parameter open model from Alibaba's Qwen team, used about 200 million output tokens to complete the Artificial Analysis Intelligence Index, against a median of 82 million for comparable open-weights models, according to the independent evaluator's model page. The same page reports 40.4 output tokens per second, about half the 82.6 median it cites for open-weights peers.
Those figures sit beside a vendor model card that is mostly about gains. The card is on Hugging Face, lists the Apache-2.0 licence, shows no download gate, and dates the release to August 2026.
What the card says it is
Qwen3.8-27B is a dense causal language model with a vision encoder, so it takes images and video. It has 64 layers, a hidden dimension of 5,120 and a feed-forward dimension of 17,408. Layers alternate Gated DeltaNet blocks with Gated Attention blocks in a 3:1 pattern, and the model uses multi-token prediction, a training technique that predicts several tokens at once.
The native context window is 262,144 tokens, extensible to 1 million with YaRN scaling. The card warns that YaRN can reduce quality on shorter texts. Thinking is on by default and can be switched off per request, with a reasoning_effort setting of xhigh, medium or low.
The card gives no minimum GPU memory. Its serving examples use vLLM, SGLang and TokenSpeed, and its Docker example sets 32 GB of shared memory.
Vendor-reported gains over Qwen3.6-27B
All figures below come from Qwen's own card and are vendor-supplied. The card compares the model against its predecessor, Qwen3.6-27B, and against Claude Opus 4.6 Max on some rows.
| Benchmark (vendor-reported) | 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 |
| HLE | 30.8 | 24.0 |
Opus 4.6 Max still scores higher on some rows, the card shows, including 78.2 on Terminal Bench 2.1 and 40.0 on HLE. The card also lists 89.2 on GPQA Diamond and 90.3 on LiveCodeBench v6.
What Artificial Analysis measured
Artificial Analysis scores the model at 34 on its Intelligence Index, in its xhigh reasoning setting. The page cites a median of 8 for comparable open-weights models. It prices the model at $0.50 per million input tokens and $3.00 per million output tokens on Alibaba's own API, and puts the cost to run each Intelligence Index task at $1.01, ranked 20th of 23.
- Qwen3.8-27B200 million tokens
- Median, comparable open-weights models82 million tokens
Source: Artificial Analysis model page for Qwen3.8 27B (xhigh), accessed 2026-10-23
The pairing explains the cost rank. A model that talks for 200 million tokens to finish the test pays for every one of them at $3.00 per million, and a reader waiting on it sees 40.4 tokens a second. The Terminal's report on Gemini 4 Argon pricing and the Decisions API report show hosted rates that look cheaper per token. Per-token price alone says little when output length differs this much.
What is not known
Artificial Analysis tested the xhigh setting. The page covered only that setting, so the token count at medium or low effort is not available from it. A user who turns thinking off, as the card allows, may see a very different profile.
No evaluator in the sources reviewed has reproduced the card's 61.7 on SWE-bench Pro. The Qwen team has not published a token-usage figure of its own. Artificial Analysis' page is the one place the 200 million figure appears, and the next check is whether the lab responds with a lower-effort setting that closes the gap.
Sources
More in A.I.
- 01Swedish Government Offices Block ChatGPT and Claude, Steer Staff to CopilotThe block starts Monday, Oct. 26, on work computers at Regeringskansliet, after the Tax Agency and the National Board of Institutional Care made similar moves.
- 02OpenAI's Reported $30 Billion Raise Now Prices the Company at $1.25 Trillion, Down From $1.4 TrillionNothing is signed and OpenAI declined to comment. The valuation in the reporting has moved three times in a month, and the March round is the only figure OpenAI has confirmed.
- 03Gemini 4 Argon's $2 Introductory Price Doubles on November 30Google's launch post and a newsletter written days earlier disagree on who can use the model, and the independent scores sit alongside a hallucination figure that only one of them reports.
- 04Berget AI Says a Texas Request Emits 8.5 Times More CO2 Than Sweden's. The Grid Gap Is 42 Times.The Swedish inference host's own calculator attributes 89% of a Swedish request's footprint to building the GPUs, which is why its headline ratio is a fifth of what grid data alone would give.