Qwen3.8-27B is now available on GraphN

Published: August 14, 2026
Reading time: 4 min

The new 27-billion-parameter Qwen model is live as managed inference with a 262,144-token context window, image and video understanding, thinking controls, and tool calling.

Qwen3.8-27B is now available in GraphN under the model alias qwen3.8-27b. Teams can select it in the model catalog, attach it to an agent, or call it through GraphN's OpenAI-compatible inference API.

The released model is a dense 27B vision-language model under the Apache 2.0 license. GraphN supports its full 262,144-token context window. The model card also documents an optional extension to one million tokens; that extension is not part of this launch.

What Qwen3.8-27B adds

Qwen3.8-27B combines capabilities that often require separate model choices:

Text, image, and video input for document, screenshot, diagram, and visual-agent tasks
Thinking mode enabled by default by the model, with request-level control when a direct answer is preferable
Tool calling and structured output for agents that need to act on external systems
262K context for large repositories, long documents, and extended task histories
Apache 2.0 open weights for teams evaluating portability and deployment control

The released architecture uses 64 language-model layers and alternates Gated DeltaNet linear-attention blocks with full-attention blocks. That hybrid design is intended to keep long-context work practical while preserving full-attention capacity throughout the network.

What the reported benchmarks show

Qwen's model card reports substantial gains over Qwen3.6-27B at the same parameter count. These are vendor-reported results, not GraphN benchmarks, and several use Qwen's own evaluation harnesses or corrected benchmark data. They are useful directional evidence, not a substitute for testing the model on your workload.

Scroll horizontally to compare

Workload	Qwen3.8-27B	Qwen3.6-27B
Terminal Bench 2.1	73.0	63.4
SWE-bench Pro	61.7	53.5
IFBench	79.5	69.1
OSWorld-Verified	84.3	63.9
WebArena-Verified	64.8	48.8
SWE-MM	38.6	25.7

The pattern matters more than a single score: the reported gains span terminal coding, repository work, instruction following, computer use, browser use, and multimodal software engineering. That makes Qwen3.8-27B a candidate for agents that must both reason over visual input and complete multi-step tool-driven work.

Qwen3.8-27B on GraphN

GraphN exposes the model through the same surfaces as its other built-in chat and vision models:

Open Models in a GraphN workspace.
Select Qwen3.8 27B (image+video).
Choose the model on an agent or workflow node.
Enable thinking when the task benefits from a longer reasoning path.

The public alias is qwen3.8-27b. The underlying checkpoint remains Qwen/Qwen3.8-27B, so prompts and model behavior stay aligned with the released weights.

You can also test it from the GraphN CLI:

bash

graphn model chat qwen3.8-27b \
  --thinking \
  -m "Review this implementation plan, identify the highest-risk assumption, and propose a test."

For applications, GraphN keeps the inference boundary OpenAI-compatible. Existing clients can switch the model field to qwen3.8-27b while retaining the standard chat-completions request shape.

When to choose it

Choose Qwen3.8-27B when the same agent needs several of these at once:

repository-level coding plus tool use
screenshots, diagrams, documents, or video alongside text
long context without moving to the largest model tier
controllable thinking for a mix of fast and difficult requests
an open-weight model with a permissive license

The smaller Qwen3.5 9B vision model remains the default for lighter multimodal tasks. Qwen3.8-27B is the higher-capability option for harder coding, agentic, and visual workloads. As with any model change, evaluate it against representative tasks, latency targets, and failure cases before moving a production workflow.

We put the launch claim to work

Qwen's launch table compares this model against Claude Opus 4.6 and reports mixed results on document benchmarks. We tested that claim ourselves with a production document workflow and both models on the same pages: Qwen3.8-27B vs Claude Opus 4.6 on real document work. Short version: parity on accuracy, about one fifth the cost.

Start with a real task

The fastest useful evaluation is not a generic chat prompt. Give Qwen3.8-27B one task that combines the capabilities you expect to use in production: a repository plus an issue, a screenshot plus a browser tool, or a long document plus a structured-output contract.

If you have a Qwen fine-tune or another checkpoint of your own, see how to import custom Hugging Face or S3 weights without managing GPUs. In our Qwen3.8-27B import test, a cached 55.6 GB checkpoint prepared in 15 seconds, and a fully warm 131K-context, same-node deployment returned to Running in 2m30s.

Explore managed inference on GraphN, review the models guide, or select qwen3.8-27b in your workspace to begin.

Primary sources

Qwen3.8-27B model card
Defines the 27B dense vision-language architecture, thinking controls, 262,144-token context, and Qwen-reported benchmark results.
Qwen3.8-27B configuration
Confirms the 262,144 maximum position setting and image and video token support in the released checkpoint.
Qwen3.8-27B license
Confirms that Qwen3.8-27B is released under the Apache License 2.0.
Qwen3.8 vLLM recipe
Documents the recommended vLLM serving path for the released model.
GraphN models reference
Documents how built-in chat and vision models are selected in GraphN agents and workflows.