Qwen3.8-27B on Artificial Analysis: No Index Yet, and the Data That Stands In

Artificial Analysis has not indexed Qwen3.8 27B. Checked 2026-08-15, the model's page on artificialanalysis.ai returns a 404 — no Intelligence Index, no output speed, no AA price, no independent benchmark. The Apache 2.0 weights shipped on August 13–14, so that will probably change within days or weeks; it has not changed yet. Until it does, every Qwen3.8 27B number you can find is Ali​baba's own model card or a news outlet repeating it. Here is the one-line answer, the AA data that does exist for the Qwen3.8 family, the official specs, and an honest read of what the "it beats everything" headlines actually are.

Sourcing note: everything below is labeled. Official means taken from Ali​baba's Qwen/Qwen3.8-27B model card on Hugging Face (August 2026). Independent means measured and reproduced by an outside party — nothing for the 27B qualifies yet. Relayed means a news outlet copying vendor numbers, which is not independent.

The one-line answer, dated

• Qwen3.8 27B on Artificial Analysis: no page, no score. The URL artificialanalysis.ai/models/qwen-3-8-27b returned "404 Page not found" when checked 2026-08-15.

• The "Qwe​n 3.8" AA scores circulating online — Intelligence 58, Agentic Index #1 — belong to Qwen3.8-Max, the 2.4T-parameter API flagship. Not the 27B.

• The most recent 27B AA does track is the predecessor Qwen3.6-27B: Intelligence Index ~37–38 in reasoning mode, ~55–59 tokens/sec output. That is your baseline for what an indexed 27B looks like.

What Artificial Analysis actually shows for the Qwen3.8 family

AA's leaderboard lists "Qwen3.8 2.4T A95B" — that is Qwen3.8-Max — at Intelligence Index 58, the second-best open-weights model on the index (Artificial Analysis models page, accessed 2026-08-15). On August 6, 2026, AA's Agentic Index ranked Qwen3.8-Max #1 globally — the first time a Chinese lab topped it, ahead of Claude Opus 5 and GPT-5.6 — as widely reported from AA's leaderboard that day.

None of those numbers transfer to the 27B. Qwen3.8-Max is a 2.4T-parameter mixture-of-experts API model with roughly 95B active parameters; the 27B is a dense, self-hostable download. They share a generation label, not an evaluation.

The predecessor is your baseline: Qwen3.6-27B on AA

AA has tracked Qwen3.6-27B for months. Depending on snapshot it scores ~37–38 on the Intelligence Index in reasoning mode (around 31 in non-reasoning mode) at ~55–59 tokens/sec — a single-GPU model holding its own against mid-tier frontier models from a year ago. That is the shape of data to expect when AA indexes the 3.8-27B: an index in the same ballpark, a speed on consumer hardware, and a price that mostly reflects the hosting choice. Treat it as a baseline, not a prediction — the 3.8-27B adds hybrid attention and native video input, either of which could move the numbers.

The official data: Ali​baba's model card, not independent

These are the verified specs from the Qwen/Qwen3.8-27B model card (Apache 2.0, August 2026):

• 27B dense parameters (28B with the vision encoder), 64 layers, hidden size 5,120, vocabulary 248,320.

• Hybrid attention: 48 linear-attention heads (Gated DeltaNet) plus 16 full Gated Attention heads — roughly a 3:1 linear-to-full split.

• Context: 262,144 tokens native, extensible to 1,000,000 via YaRN.

• Input: text, image, and video natively; thinking mode on by default.

• Size: ~55.6 GB in BF16; ~16–17 GB of weights at Q4 — a 24 GB GPU is the practical single-card target.

Vendor-reported benchmarks from the same card (Ali​baba's word, unreproduced): Terminal-Bench 2.1 73.0; SWE-bench Pro 61.7; LiveCodeBench v6 90.3; GPQA Diamond 89.2; OSWorld-Verified 84.3.

What the "it beats everything" headlines actually are

A Chinese tech-news relay (Gate.it, mid-August 2026) leads with "Qwen3.8 27B beats Meta's Muse Glimmer across all 8 benchmarks" — Terminal-Bench 2.1 73.0 vs 51.7, SWE-bench Pro 61.7 vs 51.2 — and claims wins over Claude Opus 4.6 in 15 of 19 overlapping tests. Those figures are Ali​baba's model-card numbers copied into a news summary. No outside lab has run the 27B, and the relay does not claim one did. "Third-party coverage" here means third-party repetition, not third-party verification.

Why there is no AA price for the 27B — and what that tells you

AA prices models by API endpoint. The 27B has no official hosted API yet: Ali​baba's Qwe​n Cloud hosting is "coming soon" with a 1M-token context. There is nothing to price. That is the real contrast at the center of this model. Qwen3.8 27B is an open-weights download; its price is your electricity bill. At Q4 on a 24 GB card it costs $0 per token. The API-only flagship Qwen3.8-Max is $2 per million input / $6 per million output. A self-hosted model does not appear in an API-pricing index precisely because it is free per token — the opposite situation from every API model AA does price.

What to use until AA indexes it

• You want an independently-scored Qwe​n 27B today: run or call Qwen3.6-27B. It is AA-indexed (Intelligence ~37–38 in reasoning mode), Apache 2.0, self-hostable, and it is the model the 3.8-27B has to beat. If you would rather call it over an API than run it, it is in the OrcaRouter catalog at provider list price with 0% markup — one key, no second contract.

• You want the API flagship: Qwen3.8-Max is AA-indexed at 58 and Agentic #1, at $2/$6 per 1M tokens. You are paying for a 2.4T MoE, not a local model.

• You want the 3.8-27B specifically: download the Apache-2.0 weights and self-host. Treat every benchmark as Ali​baba's word until an independent harness publishes. Watch the AA model page (artificialanalysis.ai/models/qwen-3-8-27b) and the open harnesses — when a real number lands, it will be obvious.

The bottom line

Qwen3.8 27B has no Artificial Analysis score because Artificial Analysis has not measured it — and no AA price because it is a self-host model, not an API. Every "Qwe​n 3.8 beats X" headline you can find today is Ali​baba's model card in a news costume. The numbers that are genuinely independent in this story belong to the siblings: Qwen3.8-Max at Intelligence 58 with the #1 Agentic Index, and Qwen3.6-27B at ~37–38. Anchor on those; treat the 27B's own figures as vendor claims until someone independent runs it. Check back — this page will change the day AA publishes.