Qwen3.8-Flash-Next

2026-08-26 · Publisher : Alibaba · Domain : Artificial Intelligence

Alibaba releases Qwen3.8-Flash-Next, an open-weights multimodal model with 125 billion parameters, 6 billion of them active, presented as the experimental preview of the architecture of the future Qwen4 generation.

### Context After Qwen3.5 and Qwen3.6, all built on the combination of Gated DeltaNet and Gated Attention, the Qwen team releases a new architecture for the first time before the final model. The vendor raises the question of scaling through efficiency rather than size alone. ### What the model brings - Qwen Sparse Attention (QSA): in place of Gated Attention, a sparse attention that operates on micro-blocks rather than token by token, to reduce latency on long context. - Gated Residual: widened residual streams modulated by a data-dependent read gate and a scalar write gate per branch. - N-gram Embedding: 51B embedding parameters indexed by bigrams and trigrams (20 million entries), easy to offload from GPU memory. - Training: Muon and AdamW optimizers depending on the weights, without batch-size ramp-up (vendor announcement). - A technical report is published in the GitHub repository QwenLM/Qwen3.8-Flash-Next. ### Architecture and specifications - 125B main parameters, 6B activated, plus 51B of n-gram embedding and a 4B MTP head; 48 layers; layout 12 x (3 x (Gated DeltaNet then MoE) then 1 x (QSA then MoE)); hidden dimension 2560. - 512 experts, 10 routed plus 1 shared (third-party sources); context 262,144 tokens natively, up to 1M. - Modalities: text, image, video. - License: Qwen Community License 1.0, MIT-style with conditions. Commercial use is free with display of the model name beyond 100 M monthly users or USD 20 M of monthly revenue, and a separate license from Qwen is required for any inference service (Model as a Service) or AI work assistant, with no revenue threshold; internal use exempt. - Weights: Hugging Face, Qwen/Qwen3.8-Flash-Next (and FP8). Knowledge cutoff: not disclosed. ### Cost and access | Item | Value | |---|---| | Hosted version Qwen3.8-Flash | 1M of context by default, built-in tools; pricing not disclosed in the source consulted | | Self-hosting | Compatible with Transformers, vLLM, SGLang; minimum hardware not disclosed | ### Benchmarks | Test | Qwen3.8-Flash-Next | Qwen3.8-27B | Qwen3.7-Plus (closed) | DeepSeek-V4-Flash-0731 | Opus 4.6 Max (closed) | |---|---|---|---|---|---| | SWE-bench Pro | 62.5 | 61.7 | 55.8 | 56.0 | 53.4 | | DeepSWE 1.1 | 58.7 | 42.2 | 16.5 | 54.4 | -- | | GPQA Diamond | 91.7 | 89.2 | 90.3 | 90.8 | 91.3 | | HLE | 35.9 | 30.8 | 34.7 | 33.8 | 40.0 | | LiveCodeBench v6 | 91.9 | 90.3 | 89.6 | 90.6 | 88.8 | | IFBench | 81.3 | 79.5 | 79.1 | 79.2 | 62.5 | Source: Hugging Face model card for Qwen3.8-Flash-Next, scores published by the vendor. ### Versus closed models Against Opus 4.6 Max and Qwen3.7-Plus, both closed, Qwen3.8-Flash-Next is ahead on SWE-bench Pro (62.5 vs 53.4 and 55.8), LiveCodeBench v6 (91.9 vs 88.8 and 89.6) and GPQA Diamond against Qwen3.7-Plus (91.7 vs 90.3), and slightly against Opus 4.6 Max (91.3). It remains behind Opus 4.6 Max on HLE (35.9 vs 40.0). With 6B active parameters, it claims performance superior to the dense 27B of the same family on the tests in the table, which is the vendor's message on efficiency. API prices are not disclosed, so no price gap is quantified. The Community 1.0 license is more restrictive than the Apache 2.0 of Qwen3.8-27B: worth keeping in mind before any inference offering. ### Key takeaways The model serves to examine the Qwen4 architecture in public before its release, with results already competitive against closed models of the previous generation. It is not under Apache 2.0.

Tags : LLM, Poids ouverts, Alibaba, Qwen, MoE, Qwen Community License 1.0, Architecture