Qwen3.8-Flash: Million-Token Intelligence, Built for Agentic Work

Qwen3.8-Flash is now available on RouterHub as qwen/qwen3.8-flash. The model combines a 1-million-token context window, text, image, and video input, up to 128K output tokens, and built-in tool support. Its published evaluation profile is strongest across coding, agentic workflows, professional knowledge work, and general reasoning.

Qwen3.8-Flash at a Glance

Item	Published specification
Provider	Qwen, Alibaba Group
RouterHub model ID	qwen/qwen3.8-flash
Input	Text, image, and video
Output	Text
Context window	1 million tokens
Maximum output	128K tokens
Published base architecture	125B-parameter main model, plus 51B N-gram embedding parameters
Activated parameters	6B per token
Production features	Built-in tools and API compatibility for OpenAI- and Anthropic-style integrations

Qwen describes Qwen3.8-Flash as the production version built on Qwen3.8-Flash-Next. The production model extends the underlying release with a 1-million-token context window by default and official built-in tools.

Published Coding and Agentic Results

Qwen’s published evaluations for Qwen3.8-Flash-Next, the model underlying Qwen3.8-Flash, show a broad performance profile across repository-level engineering, tool use, and long-horizon tasks.

Benchmark	Published result	What it evaluates
DeepSWE 1.1	58.7	Agentic software engineering
SWE-bench Pro	62.5	Difficult repository-level software engineering
SWE-bench Multilingual	81.0	Multilingual software engineering
NL2Repo-Bench	48.1	Repository-level code generation
CoWorkBench	73.9	Long-horizon office and professional work
JobBench	55.7	Professional job tasks
Toolathlon Verified	73.5	Real-world tool use

The results point to a model designed for more than short code completion. Qwen3.8-Flash is positioned for tasks that require reading a large working set, acting through tools, and sustaining progress across multiple steps.

Published Reasoning and Instruction-Following Results

Benchmark	Published result	What it evaluates
IFBench	81.3	Instruction following
GPQA Diamond	91.7	Graduate-level scientific reasoning
HLE	35.9	Difficult multidisciplinary reasoning
LiveCodeBench v6	91.9	Competitive coding

These benchmarks use different task sets, scales, and evaluation harnesses. The results were published by Qwen and have not been independently reproduced by RouterHub.

Long-Context Efficiency

Qwen3.8-Flash-Next uses a hybrid Gated DeltaNet and Qwen Sparse Attention design. At a 1-million-token context length, Qwen reports that the sparse-attention kernel reached up to 7.6× faster prefill and 4.9× faster decode in its experiments. In an online-serving setup with a 90% prefix-cache hit rate, Qwen also reports 8.6× the prefill throughput of Qwen3.7-Plus.

Those figures matter most for applications that work with large repositories, long documents, extended conversations, or repeated shared context. They indicate that the model’s long-context capability is paired with an architecture designed to retrieve relevant information without applying full attention to every token at every layer.

Where the Performance Profile Stands Out

Qwen3.8-Flash is a strong candidate for evaluation when a workload combines several of the following requirements:

• repository-scale coding and debugging;
• long-horizon agentic execution;
• tool use across multi-step workflows;
• professional document and office tasks;
• multimodal analysis across text, images, or video;
• long-context reasoning with large working sets.

The published results provide a useful starting point, but production evaluation should still use representative prompts, tools, context sizes, and success criteria from the intended application.