Qwen3.8 Flash Next
Open-weight experimental preview of the Qwen4 architecture: hybrid-attention MoE (125B total, 6B active) with vision encoder for coding, agent tasks, and image and video understanding
- Strongest at
- Writing Top 1% of the past year's models · one benchmark so far
- Context window
- 262.1K tokens in · 131.1K out
Capability profile
Scored against 91 models released in the past year. 50 is the average model; further out is better.
No results yet: Science, Visual reasoning, Computer use
Where these scores come from
Each capability combines the published benchmarks listed here, rescaled so 50 is the reference average and 15 points is one standard deviation. A hollow dot rests on a single benchmark.
- Coding LiveBench: Coding + Agentic Coding
- Writing LiveBench: Instruction Following
- Reading & documents LiveBench: Language + Data Analysis, ExtractBench: document extraction
- Reasoning LiveBench: Reasoning
- Math LiveBench: Mathematics
Qwen's past year
48 release days in the past year, about one every 8 days. The next is expected around Nov 2, 2026. Forecasts ↗
Specifications
Specifications and prices from Models.dev; availability and limits may differ by provider. Benchmarks from Epoch AI (CC BY 4.0), LiveBench and the other sources named above.