Qwen3.8 27B: A Dense Apache-Licensed Model With Agent Ambitions

A dense 27B model is supposed to be the sensible, modest option: good enough for routine work, small enough for one GPU, never in the conversation about frontier capability. Qwen3.8 27B, released in mid-August 2026, comes with vendor numbers that put it in that conversation. Whether they hold on your work is the thing to find out.

Its other notable property is less glamorous and more useful. It ships under Apache 2.0, accepts images and video, and is served by a long list of providers at wildly different prices and precisions. That spread is where most of the practical decisions are.

What changed from Qwen 3.6 27B

Alibaba describes it as a 27B dense native vision-language model that improves on Qwen 3.6 27B in coding and office tasks across text and visual input. The model card shows an architecture that looks little like the classic dense transformer:

• 64 layers in 16 blocks, each made of three Gated DeltaNet layers and one gated full-attention layer.
• Multi-token prediction, trained with multiple steps.
• A vision encoder; the Hub lists the full checkpoint at 28B parameters, which is the language model plus that encoder.
• Context: 262,144 tokens natively, extendable to 1M with YaRN scaling. The card warns that enabling it can hurt performance on shorter texts, so switch it on only for genuinely long inputs.

The Gated DeltaNet layers use linear attention, which keeps a fixed-size state rather than a key-value cache that grows with every token. Only a quarter of the layers keep a conventional cache, which makes long contexts much cheaper in memory than the same parameter count would suggest. The predecessor is covered in the Qwen 3.6 guide.

The vendor numbers, read carefully

The card reports, among others:

• SWE-bench Pro 61.7 and Terminal-Bench 2.1 73.0
• LiveCodeBench v6 90.3
• GPQA Diamond 89.2 and Humanity's Last Exam 30.8
• IFBench 79.5
• OSWorld-Verified 84.3 and AndroidWorld 81.9 for screen-operation tasks

For context, Z.ai reports 62.1 on SWE-bench Pro for GLM-5.2, a mixture-of-experts model of roughly 744B total parameters. A dense 27B reporting within half a point of that would be a notable result if it holds up.

It is also a vendor figure on a vendor scaffold, and those routinely run ahead of independent harnesses. It cannot be compared with the 77.2 often quoted for Qwen 3.6 27B either, because that was SWE-bench Verified, a different and easier benchmark. Benchmark contamination explained covers another reason to discount any single headline number.

Same weights, very different endpoints

OpenRouter lists around twenty providers for this model, and they are not interchangeable. At the time of writing:

• Precision ranges from FP4 through FP8 to BF16 and FP16.
• Context is 262,144 tokens on most endpoints, 1M on a few, including Alibaba's, and 65,536 on one.
• Input prices run from about $0.02 to about $0.99 per million tokens, and output from about $0.50 to $3.20.

The cheapest endpoint is not necessarily the same model in any practical sense. A 4-bit serving of a reasoning model can lose accuracy on exactly the long, careful tasks you picked it for. Quantised model quality loss covers how to check. If results matter, pin a provider whose precision you know, and test on that endpoint rather than on whichever one the router chose that day.

There is also a :free variant on OpenRouter with a 262,144-token window. It is rate-limited and suited to trying the model, not to production; free tier AI limits compared covers what to expect.

The thinking controls

Thinking is on by default and can be turned off per request with enable_thinking set to false. When it is on, reasoning_effort accepts xhigh, the default, medium or low. A preserve_thinking option, also on by default, keeps reasoning from earlier turns in the history.

Those defaults favour quality over cost. Maximum effort writes the most thinking tokens, preserved reasoning lengthens every later prompt, and on hosted endpoints output is the expensive side of the bill. For classification, extraction or short edits, turn thinking off or drop to low and measure whether anything gets worse. The card recommends temperature 1.0, top_p 0.95 and top_k 20 with thinking on, and a lower temperature of 0.7 with top_p 0.8 without it.

Running it yourself

The arithmetic is friendly. At 28B parameters, BF16 weights take roughly 56 GB, which needs an 80 GB-class card. Qwen publishes an official FP8 checkpoint at roughly 28 GB, which fits a 48 GB card with room for context. A 4-bit quantisation, at roughly 14 to 16 GB, brings it within reach of a 24 GB consumer card, with the quality caveats above.

Context memory is where the linear-attention layers pay off: only 16 of the 64 layers keep a growing cache. The card recommends SGLang, vLLM or TokenSpeed for serving. Which open model fits your GPU works through the memory budget, and single-GPU models compared places it against the alternatives.

27B or the 2.4T sibling

The 3.8 generation's other open model is the Qwen3.8 2.4T A95B, the open version of the flagship. The choice is rarely close:

• Pick the 27B if you want one GPU, image or video input, the option to turn thinking off, and an Apache 2.0 licence with no thresholds.
• Pick the 2.4T if you need the most capable open Qwen, work only with text, can run a multi-node cluster, and have read its custom licence.
• Test both through hosted endpoints first. Twenty of your own tasks will tell you whether the gap is worth close to ninety times the memory.