Qwen3.8-Flash

How it is built

Two names cover one model. The open weights are Qwen3.8-Flash-Next, and Qwen3.8-Flash is the API version built on them.¹¹1Qwen, Qwen3.8-Flash-Next model card, August 2026. SWE-bench Pro with the Claude Code harness, temperature 1.0, top_p 0.95 and a 256K context window. https://huggingface.co/Qwen/Qwen3.8-Flash-Next

The main model has 125B parameters, with 6B active for each token.¹¹ Qwen counts two further parts apart: an n-gram embedding of 51B parameters and a 4B module for multi-token prediction.¹¹

In every layer a token meets 10 routed experts out of 512, plus 1 shared expert.¹¹ The 48 layers repeat a block of three Gated DeltaNet layers and one layer of Qwen Sparse Attention (QSA), 12 times over.¹¹ Hidden states are 2,560 wide, and QSA layers use 24 query heads with 2 key-value heads.¹¹ A vision encoder brings images and video into the same model.¹¹

What is new

Qwen presents the design as an experimental preview of the architecture planned for Qwen4.¹¹ The weights and the blog post came out on 26 August 2026.²²2Qwen, Qwen3.8-Flash-Next blog post, 26 August 2026. Training "takes only about 1/9 as much" as Qwen3.7-Plus. https://qwen.ai/blog?id=qwen3.8-flash-next

QSA replaces the Gated Attention of earlier Qwen models. A small indexer picks context in micro-blocks rather than single tokens, with a budget of 512 blocks, or 2,048 tokens.¹¹ Gated Residual widens the residual stream into 4 branches. An element-wise gate that depends on the data controls what is read, and each branch has a scalar gate for what is written.¹¹

By Qwen's account, training cost about a ninth of what Qwen3.7-Plus took.²² On SWE-bench Pro the card gives it 62.5. The dense Qwen3.8-27B scores 61.7 in the same table, less than a point behind, and Qwen3.7-Plus 55.8.¹¹ All three ran in Claude Code with a 256K context.¹¹

Running it

The Qwen Community License 1.0 has no revenue threshold, unlike the licence of Qwen3.8-Max. Any company that runs a model-as-a-service or AI work assistant business needs a separate licence for commercial use, whatever its size.³³3Qwen, Qwen Community License 1.0, condition 2. https://huggingface.co/Qwen/Qwen3.8-Flash-Next/blob/main/LICENSE

Qwen says the n-gram table is easier to offload than experts, which suits accelerators with little memory.¹¹ Past its native 262,144 tokens, the context stretches to 1,000,000 with YaRN.¹¹ A request can turn thinking off.¹¹

On QwenCloud the model runs as qwen3.8-flash, with 1M tokens of context and up to 131K of output.⁴⁴4QwenCloud, Qwen3.8-Flash model page, prices per million tokens as listed on 29 September 2026. https://www.qwencloud.com/models/qwen3.8-flash It costs $0.15 per million input tokens and $0.47 per million output tokens. Cached input costs $0.016.⁴⁴