Alibaba's Qwen team has released the multimodal MoE model Qwen3.8-Flash and open-sourced its early preview Qwen3.8-Flash-Next, which serves as a preview of the Qwen4 architecture. The new model achieves significant efficiency gains, outperforming the previous flagship Qwen3.7-Plus at just 1/9 of the training cost, with particularly strong results in coding and office tasks.
SynthePulse Insight · AI deep reading
Version 1 · 1 source
Alibaba's Qwen releases multimodal MoE model Qwen3.8-Flash and open-sources its Next version, which the company defines as an early preview of the Qwen4 architecture. In efficiency, stability, and agent benchmarks, the new model achieves multiple leads at a fraction of the training cost of the previous flagship, though it still trails on some benchmarks, and the open-source version differs from the production API version.
On August 26, Alibaba's Qwen team officially released the multimodal MoE model Qwen3.8-Flash and simultaneously open-sourced its Next version, Qwen3.8-Flash-Next. The company defines Flash-Next as an early preview of the Qwen4 architecture, with the early open-source release intended to allow inference frameworks and infrastructure developers such as vLLM and SGLang to adapt before the official Qwen4 family launch.
The production version, Qwen3.8-Flash, will soon be available via the QwenCloud API, priced at $0.16 per million input tokens and $0.47 per million output tokens. The model has 125B parameters and includes a 51B N-gram component.
According to the official report, compared to the previous flagship Qwen3.7-Plus (397B/A17B), Flash-Next uses only 1/3 activated parameters, 1/3 training tokens, and about 1/9 training FLOPs, achieving leads on 8 of 14 pretraining benchmarks, with the rest trailing by at most 2.6 points. This efficiency improvement is the core highlight of this release.
In terms of stability, the report provides a reproducible stress-test method: doubling the constant learning rate to simulate large-scale training instability. At 4x the optimal learning rate, the old architecture (AdamW + Qwen3.5 structure) frequently exhibits loss spikes, while the new recipe (Muon + gated residual) achieves zero spikes. Additionally, production training uses no explicit clipping techniques like qk-clip.
The official post-training benchmark comparisons include Qwen3.7-Plus, DeepSeek-V4-Flash, Claude Opus 4.6 Max, and others. In agentic coding, DeepSWE 1.1 scores 58.7, while Qwen3.7-Plus scores only 16.5 and DeepSeek-V4-Flash 54.4; SWE-bench Pro scores 62.5, and SWE-bench multilingual version 81.0.
In long-horizon office/agent benchmarks, CoWorkBench 73.9, JobBench 55.7, and Toolathlon 73.5, all clearly leading comparison models. In general benchmarks, GPQA Diamond 91.7, LiveCodeBench v6 91.9, and IFBench 81.3. In multimodal benchmarks, AndroidWorld 84.5, MathVision with code interpreter 95.7, OSWorld 2.0 partial 52.3, and long video LVBench 76.6.
This report's information primarily comes from Qwen's official X account and reposts, which are first-party releases, but all benchmark data are official self-tests and have not been independently verified by third parties. Some data (such as training cost comparisons) come from official reports and should be considered source claims.
Qwen3.8-Flash-Next achieves performance close to or exceeding the previous flagship at a very low training cost, and excels in agentic and multimodal benchmarks, while demonstrating the new architecture's stability advantages. However, note that the open-source version differs from the production API version, and all benchmarks are official self-tests, so actual performance awaits third-party verification.
Secondary coverage
This article also covers this event; its overall headline is not the event headline.