# Alibaba Puts Max-Class Weights on Hugging Face as Qwen3.8-27B Runs Agent Benchmarks on One GPU

The 27-billion-parameter model reports 73.0 on Terminal-Bench 2.1, up from 63.4 for its predecessor, and the 2.4-trillion-parameter mixture-of-experts flagship, previously available only through paid access, is now downloadable in Alibaba's Max line.

- Published: 2026-08-17T06:18:43.890Z
- Canonical: https://polylog.news/ai/2026-08-17/alibaba-puts-max-class-weights-on-hugging-face-as-qwen3-8-27
- Publisher: Polylog (AI desk)
- Section: tech
- Sources: [Simon Willison](https://simonwillison.net/2026/Aug/16/qwen-38-27b/), [Latent Space (AINews)](https://www.latent.space/p/ainews-qwen-38-max24t-and-27b-new), [Hugging Face](https://huggingface.co/Qwen/Qwen3.8-2.4T-A95B)

Alibaba's Qwen team released two open-weight models this weekend that change the cost calculation for teams deciding whether to rent frontier capability or host it themselves. Qwen3.8-27B is a 27.78-billion-parameter model that accepts text, images and video, ships under the Apache 2.0 license, and carries a native context window of 262,144 tokens, extendable toward one million tokens using a context-extension method (YaRN scaling). Alongside it, Alibaba published [Qwen3.8-2.4T-A95B](https://huggingface.co/Qwen/Qwen3.8-2.4T-A95B), a mixture-of-experts model with 2.4 trillion total parameters and roughly 95 billion active per forward pass. Alibaba describes it as the first Max-class Qwen checkpoint released with downloadable weights.

The figures come from Qwen's own model card and have not been independently reproduced. On that card, the 27-billion-parameter model reports Terminal-Bench 2.1 rising to 73.0 from 63.4 for Qwen3.6-27B, DeepSWE 1.1 rising to 42.2 from 13.3, OSWorld-Verified rising to 84.3 from 63.9, and SWE-MM rising to 38.6 from 25.7. It also lists 61.7% on SWE-Bench Pro, 89.2 on GPQA Diamond and 90.3 on LiveCodeBench. The larger 2.4-trillion-parameter model is reported at 93.0 on PaperBench and 86.6 on Terminal Bench 2.1, [comparable to closed models from Anthropic and OpenAI on several agentic tests](https://www.mindstudio.ai/blog/qwen3-8-2-4t-a95b-benchmarks) while trailing them on others.

The most useful independent assessment so far comes from developer Simon Willison, who [tested the 27B model locally](https://simonwillison.net/2026/Aug/16/qwen-38-27b/) and called it excellent, while noting that it defaults to long reasoning traces even on trivial prompts. That default has a direct cost consequence. A model that spends thousands of reasoning tokens on a one-line question reduces the cost advantage of running it yourself, because latency and memory use scale with the number of tokens generated, not with parameter count.

Two claims here are verified and two are not. Verified: the weights are downloadable, the license is permissive, and the 27-billion-parameter checkpoint fits on a single high-end accelerator. Not verified: whether the reported agentic scores hold up under third-party evaluation, and whether the jump on DeepSWE 1.1 from 13.3 to 42.2 reflects broad capability gains or optimization targeted at that specific test.

## What this means

A Max-class model with downloadable weights removes the last structural argument for paying frontier API rates on agentic coding work that a team could instead run on owned or rented hardware. The exposed parties are closed-model vendors that earn inference margin on mid-tier reasoning tasks. The beneficiaries are inference hosts, GPU renters and enterprises in jurisdictions that cannot buy American frontier model access. The countervailing force is the excessive reasoning behavior Willison identified. If the model keeps generating long reasoning traces by default, the actual serving cost per completed task narrows the savings that the advertised price implies.

## What to watch

- Whether independent evaluators reproduce the Terminal-Bench 2.1 and OSWorld-Verified figures, which would separate a real capability step from harness-specific tuning.
- Pricing moves by closed-model vendors on their mid-tier coding endpoints, the fastest signal that a downloadable competitor is taking share.
- How quickly quantized community builds appear and what accuracy they retain, since that determines whether the model reaches single-workstation deployments rather than data-center GPUs only.