Qwen released Qwen3.8-27B on August 14 as a dense, open-weight model built for coding and long-running agent tasks. It has 27 billion parameters, native image and video understanding, a 262,144-token context window, thinking enabled by default, and an Apache 2.0 license.
The release is less about a new chat experience and more about making a 27B model finish work inside an agent harness. Qwen reports stronger planning, better use of feedback from tools and environments, and higher completion rates across terminal, software engineering, browser, and computer-use benchmarks.
The weights and full model card are available on the official Qwen3.8-27B page.
What Qwen3.8-27B adds for agents
Stronger multi-step execution
Qwen says the model was post-trained to plan more reliably and respond better to feedback from its environment. In practice, that is the loop an agent needs: choose an action, inspect the result, correct the plan, and continue until the task is complete.
The benchmark gains support that claim within Qwen's own evaluation setup. Terminal Bench 2.1 rises from 63.4 on Qwen3.6-27B to 73.0. SWE-bench Pro moves from 53.5 to 61.7. CoWorkBench, which covers long professional tasks, goes from 61.0 to 70.7.
Reasoning depth can be set per request
Thinking mode is enabled by default. Developers can choose three reasoning-effort levels: "xhigh," "medium," and "low." Thinking can also be disabled for requests that need a direct answer instead of a longer reasoning trace.
This matters in an agent because the best setting depends on the job. A repository-wide fix can justify deeper reasoning. A short formatting step usually cannot. Qwen also warns that lower effort may reduce the time of one response while increasing total work if the agent makes more mistakes and retries.
Earlier reasoning can remain in context
The new preserve_thinking option keeps reasoning blocks from previous messages in the conversation. The model can reuse earlier decisions instead of reconstructing the plan on every turn.
For long tasks, this targets two common failures: contradicting a decision made several steps earlier and spending tokens solving the same subproblem again. The feature is enabled by default and can be turned off when an application wants to retain only the latest reasoning block.
Images and video are native inputs
Qwen3.8-27B is a vision-language model, not a text model with a separate image adapter exposed as another product. It accepts text, images, and video through the same conversation format.
That gives an agent direct access to screenshots, interface states, diagrams, scanned documents, charts, and video frames. Qwen also documents hour-scale video understanding, although the practical frame rate and token use depend on preprocessing and available memory.
The multimodal benchmark results are some of the largest jumps in the release. OSWorld-Verified rises by 20.4 points over Qwen3.6-27B. WebArena-Verified gains 16.0 points, while SWE-MM improves by 12.9 points.
A 262K native context window
The native context length is 262,144 tokens. Qwen documents extension to one million tokens with YaRN in supported serving frameworks.
The native number is the useful baseline for local deployment. Extending context is not free: the cache, input, reasoning, and final output all compete for memory. A runtime may support one million tokens while the machine can use only a fraction of that window at a practical speed.
Support for current serving and agent stacks
The official release documents deployment with Transformers, vLLM, SGLang, and TokenSpeed. Its chat template supports multi-step tool calls and tool responses, and the model card includes OpenAI-compatible API examples.
This does not mean every harness will behave identically. Tool syntax, context management, retry policy, and sampling still affect completion rates. It does mean developers can integrate the model without building around a private inference protocol.
Agent benchmark results
The following scores were published by the Qwen team. They compare Qwen3.8-27B with Qwen3.6-27B inside the release evaluation setup.
| Test | Qwen3.8-27B | Qwen3.6-27B |
|---|---|---|
| Terminal Bench 2.1 | 73.0 | 63.4 |
| SWE-bench Pro | 61.7 | 53.5 |
| CoWorkBench | 70.7 | 61.0 |
| Agents' Last Exam, Pass@1 | 20.4 | 10.6 |
| OSWorld-Verified | 84.3 | 63.9 |
| WebArena-Verified | 64.8 | 48.8 |
| SWE-MM | 38.6 | 25.7 |
| Vision2Web | 62.9 | 45.0 |
The coding results cover different levels of work. Terminal-Bench 2.1 measures tasks completed in a command-line environment. SWE-bench Pro uses long software issues that can require changes across several files. Vision2Web measures visual web development, while SWE-MM adds screenshots and other visual evidence to software engineering tasks.
The computer-use results test whether the model can act on visual interfaces. OSWorld-Verified covers desktop applications. WebArena-Verified uses reproducible websites and deterministic graders. On these two tests, Qwen3.8-27B improves by 20.4 and 16.0 points over Qwen3.6-27B.
Agents' Last Exam is still difficult. Pass@1 reaches 20.4, up from 10.6. That is close to a twofold increase, but it also means the model fails most frontier tasks on its first attempt. The result is a useful reminder that improvement and reliability are not the same thing.
How to read the numbers
These are launch results from the Qwen team, not independent measurements. Several datasets are public, but the harness and evaluation settings matter.
For SWE-bench Pro, Qwen used a Claude Code harness, a 256K context window, temperature 1.0, and top_p 0.95. The team also corrected problematic tasks and reran the baseline models. The reported 61.7 therefore should not be compared directly with a result produced on a different version or public leaderboard setup.
CoWorkBench and QwenSWEBench are internal Qwen benchmarks. They help compare models inside the release table, but independent reproduction is not yet available. The model was released today, so community evidence on long runs, quantized variants, and different agent harnesses is still limited.
The clearest conclusion is the comparison within Qwen's own table: Qwen3.8-27B improves over Qwen3.6-27B in every agent benchmark shown above, with the largest absolute gains in computer use and browser work.
What 27B means for local deployment
Qwen3.8-27B is dense. All 27 billion parameters participate in every forward pass. The language model has 64 layers arranged as a repeating 3:1 pattern of Gated DeltaNet and gated full-attention layers. It also uses multi-token prediction.
The hybrid attention design is meant to reduce the cost of long sequences without removing full attention entirely. It is the same architectural line Qwen introduced before 3.8, with the new release concentrating on capability and agent execution rather than presenting a completely different model family.
The original BF16 files occupy about 55 GB. Community GGUF quantizations range from roughly 9 GB at 2-bit to 31.5 GB at 8-bit. Common 4-bit files are around 16 to 18 GB.
Those numbers describe the model file, not the complete memory requirement. Context cache, vision processing, temporary buffers, and the operating system need additional room. The usable context length and speed will vary by quantization, runtime, and hardware.
Qwen3.8-27B will be available on Forjal in the next few days.