Alibaba pushed Qwen3.8-27B to Hugging Face about an hour ago under a plain Apache 2.0 license, and for the first time in this model generation the release arrived with its own benchmark table. Terminal-Bench 2.1: 73.0. SWE-bench Pro: 61.7. GPQA Diamond: 89.2. LiveCodeBench v6: 90.3. It is a dense 27-billion-parameter model, 64 layers, hidden dimension 5,120, with a vision encoder built into the same stack so it reads images and video natively rather than through a bolted-on adapter.

The scoreboard is the part worth pausing on. We covered the Qwen3.8-Max launch on August 3, and the story then was an absence: a 2.4-trillion-parameter flagship shipped with a price, a tagline calling it a new bar for coding, and not one evaluation number attached. That was the second Qwen3.8 event in three weeks to skip the table. This one does not skip it.

RelatedQwen3.8-Max lands at 2.4T with no benchmark table

What actually shipped tonight?

The release is two repositories, not one. There is a full-precision Qwen3.8-27B and a Qwen3.8-27B-FP8 built with fine-grained FP8 quantization at a block size of 128, which is the version most people running this on a single card will pull. Both carry the same model card and the same license. The concrete specifications:

  • 27B parameters, dense. Not a mixture of experts. Every parameter is active on every token, which is why the memory math for a single accelerator is predictable rather than a sparsity puzzle.
  • 262,144 tokens of native context, which Alibaba says extends to 1,000,000. That is flagship-tier context on a model sized for one GPU.
  • Text, image, and video in. The vision encoder is part of the architecture, and the card publishes separate vision scores: OSWorld-Verified 84.3, AndroidWorld 81.9, WebArena-Verified 64.8, OmniDocBench 1.5 at 91.1.
  • Apache 2.0. No custom community license, no acceptable-use rider, no monthly-active-user threshold. You can ship it in a commercial product without asking.
  • Three reasoning budgets. A reasoning_effort parameter takes xhigh, medium, or low, with xhigh as the default. SGLang, vLLM, and TokenSpeed are the recommended runtimes.
SWE-bench Pro scores among open-weight models Horizontal bar chart. Qwen3.8-27B leads open-weight models with a published SWE-bench Pro score at 61.7, ahead of LongCat-2.0 at 59.5, Laguna S 2.1 at 59.4, MiniMax M3 at 59.0, Kimi K2.6 at 58.6 and DeepSeek V4 Pro at 55.4. SWE-bench Pro, open-weight models with a published score vendor-reported figures · axis starts at 50 to separate a tight field Qwen3.8-27B 61.7 LongCat-2.0 59.5 Laguna S 2.1 59.4 MiniMax M3 59.0 Kimi K2.6 58.6 DeepSeek V4 Pro 55.4 genztech.blog
Fig 1 · benchmark Against the open-weight models on our AI coding leaderboard that publish a SWE-bench Pro figure, 61.7 is the highest we have on file. The axis is truncated at 50 because the field between 55 and 62 is genuinely that tight.

How does a 27B model land this close to a 2.4T flagship?

Alibaba's own numbers put Qwen3.8-Max at 86.6 on Terminal-Bench 2.1 and 67.7 on SWE-bench Pro. The 27B lands 13.6 points back on the first and 6.0 points back on the second, using roughly one percent of the parameter count and a fraction of the active compute per token. On the harder of the two coding benchmarks, the gap between Alibaba's largest model ever and a model you can run on one accelerator is six points.

Part of that is benchmark saturation at the top: SWE-bench Pro was designed to be hard, and the whole field is bunched in the fifties and sixties, so six points covers more real capability difference than it looks like. Part of it is that agentic coding scores measure the model plus its scaffolding, and a smaller model with a good harness closes distance that raw capability would not. And part of it is simply that dense mid-size models have gotten very good at the specific shape of work these benchmarks test.

Qwen3.8-27B versus Qwen3.8-Max on two coding benchmarks Grouped bar chart. On Terminal-Bench 2.1 the 27B scores 73.0 against the Max model's 86.6. On SWE-bench Pro the 27B scores 61.7 against 67.7. 27 billion dense versus 2.4 trillion sparse both figures vendor-reported by Alibaba · higher is better 27B Max (2.4T) 0 50 100 73.0 86.6 Terminal-Bench 2.1 61.7 67.7 SWE-bench Pro genztech.blog
Fig 2 · benchmark Six points on SWE-bench Pro is what separates a model sized for a single accelerator from Alibaba's 2.4-trillion-parameter flagship, on the vendor's own reporting.
 Qwen3.8-27BQwen3.8-Max
Parameters27B, all active2.4T, 95B active
ArchitectureDense + vision encoderSparse mixture of experts
LicenseApache 2.0Custom qwen3.8-max license
Context262K native, 1M extended262K native, 1.01M extended
Modalities inText, image, videoText
Terminal-Bench 2.173.086.6
SWE-bench Pro61.767.7
Realistic deploymentOne accelerator, on premisesCluster or hosted API

How much should you trust these numbers?

Not entirely, and Alibaba specifically is the reason we say that. Our AI coding leaderboard carries the widest vendor-versus-independent gap on the board against this exact maker: Qwen3.7 Max is claimed at 80.4 percent on SWE-bench Verified, and vals.ai measures it at 68.8 percent on a neutral bash-only harness. That is 11.6 points. It is not fraud, it is scaffolding. SWE-bench and Terminal-Bench score the model together with whatever agent loop wraps it, and a vendor tunes that loop for its own release.

There is a useful calibration point right next door. Alibaba published no table at all for Qwen3.8-Max on August 3, so that model sat unranked on our board until vals.ai ran it independently on August 8 and returned 85.6 percent on SWE-bench Verified, which is where it sits now at rank 11. The flagship has an outside number. The 27B does not yet.

So read 61.7 and 73.0 as best-case figures produced by the team that built the model, and expect an independent run to come in lower. How much lower is the interesting question, and it is one nobody can answer tonight. The model goes onto our leaderboard as unranked and unverified, with the vendor figures visible and clearly labelled, until an independent evaluation exists. That is the honest place for it.

Who does this actually change things for?

Anyone who needed a capable coding and agent model inside their own network. The combination on offer here is unusual: permissive license, single-accelerator footprint, a quarter-million tokens of context, and vision built in. Teams in regulated industries, or teams that simply do not want their source code leaving the building, have been choosing between a small open model that is not good enough and a hosted frontier model they are not allowed to call. A 27B that scores 61.7 on SWE-bench Pro narrows that choice considerably.

RelatedLaguna S 2.1: 8B Active Params, 70% Terminal-Bench

The vision scores matter more than they look. OSWorld-Verified at 84.3 and AndroidWorld at 81.9 are computer-use benchmarks: the model looks at a screen and drives software. Getting that in the same weights as the coding model, on hardware you own, is a different proposition from stitching together a code model and a separate VLM.

  1. Aug 3Qwen3.8-Max launches at 2.4T Priced at $2 in and $6 out per million tokens, with no benchmark table. Open weights promised for both Max and the 27B within the week.
  2. Aug 13Max weights land, 27B slips Qwen3.8-2.4T-A95B posted to Hugging Face and ModelScope. The 27B missed its window with no new date given.
  3. Aug 14Qwen3.8-27B ships with numbers Apache 2.0, full and FP8 repositories, and a complete benchmark table across coding, reasoning and vision.
  4. NextIndependent evaluation A neutral-harness run is what turns 61.7 from a claim into a ranking. Until then it stays unranked on our board.

What does it mean for the market?

Alibaba (NYSE: BABA) is now giving away, under Apache 2.0, a model that scores within six points of its own paid flagship on the hardest public coding benchmark. That is a deliberate wedge aimed at the middle of the API market, and the companies most exposed are the ones whose business is selling inference on mid-tier models rather than frontier ones. The signal for investors is not a Qwen revenue line, because there mostly is not one here. It is pricing pressure: every capable open-weight release of this size resets what customers are willing to pay for a hosted model that performs similarly, and Alibaba can absorb that because its return is cloud consumption and ecosystem position, not per-token margin. Watch whether the mid-tier API price sheets at the large providers move in the next quarter, and watch how quickly this lands in the serving stacks that enterprises actually buy.

What to watch · next 30 days
  • An independent number. If a neutral harness puts the 27B anywhere near 61.7, this is the most significant open-weights release of the month. If it lands eight points lower, it is a good model with an oversold card.
  • Real memory requirements. The card publishes no VRAM figure. FP8 at 27B suggests a single 80GB accelerator with room for context, but long-context KV cache is where that math usually breaks.
  • Whether the vision half holds up. Computer-use benchmarks are young and noisy. OSWorld-Verified at 84.3 is a strong claim that deserves scrutiny before anyone builds on it.
  • The next Qwen launch. Two releases without a table, then one with. If the table becomes the norm, this stops being a story.

Our take

We spent two posts criticising Alibaba for shipping models with no evaluation data, so it is only fair to say plainly that this release fixed the thing we complained about. A full table, across coding, reasoning and vision, on the day of release, on a permissively licensed model. That is how it should work.

The numbers themselves still need an outside referee, and this maker's track record says the gap will not be small. But there is a difference between a claim you can check and no claim at all. Tonight Alibaba published something falsifiable, which is more than it did on August 3, and considerably more than a lot of frontier labs manage.

Primary sources

Original analysis by GenZTech, built from the Qwen3.8-27B model card published by Alibaba.