Qwen3.8-Flash-Next Technical Specifications
| Specification | Qwen3.8-Flash-Next |
|---|---|
| Developer | Qwen Team, Alibaba Group |
| Model type | Multimodal causal language model with vision encoder; ultra-sparse MoE |
| Main-model parameters | 125B |
| Activated parameters | 6B per token |
| N-gram embedding parameters | 51B additional |
| MTP parameters | 4B |
| Layers | 48 |
| Hybrid attention pattern | 12 x [3 Gated DeltaNet layers + 1 Qwen Sparse Attention layer] |
| MoE experts | 512 total; 10 routed + 1 shared per token |
| Gated Residual | 4 branches; bottleneck rank 320 |
| Native context window | 262,144 tokens |
| Extended context | Up to 1,000,000 tokens with YaRN |
| Modalities | Text, image, video input; text output |
| Thinking controls | Thinking/non-thinking plus reasoning-effort controls in supported APIs |
| Deployment | Transformers, vLLM, SGLang, TokenSpeed and other supported frameworks |
| Open weights | Yes |
| CometAPI model ID | qwen3.8-flash-next |
What Is Qwen3.8-Flash-Next?
Qwen3.8-Flash-Next is an upcoming open-weight multimodal MoE model from Alibaba's Qwen team. More importantly, it is positioned as an early preview of the architecture intended to power the future Qwen4 model family, rather than simply another incremental Qwen3.8 checkpoint.
Qwen is using this release to expose its next-generation architectural direction to the open-source ecosystem before the broader Qwen4 family arrives. This gives inference-engine developers, quantization projects, model-serving frameworks, and application developers an opportunity to prepare for the architectural changes in advance.
The currently disclosed architecture introduces two important components: a GDN-based hybrid architecture and Qwen Sparse Attention (QSA). GDN refers to a gated DeltaNet-style linear-attention approach that follows the hybrid-attention direction previously explored in Qwen3-Next. QSA is presented as a new sparse-attention mechanism, although its complete technical specification has not yet been publicly documented.
Main Features of Qwen3.8-Flash-Next
- Qwen4 architecture preview: Qwen3.8-Flash-Next is explicitly positioned as an early implementation of the architectural direction that will power the upcoming Qwen4 family.
- Multimodal MoE design: The model is described as a multimodal Mixture-of-Experts model, extending Qwen's efficient sparse-model strategy toward a new generation of architecture.
- GDN hybrid architecture: The model continues the hybrid-attention research direction introduced with Qwen3-Next, combining efficient linear-attention mechanisms with conventional attention where precise retrieval is important. Qwen3-Next demonstrated that this approach can substantially improve long-context inference efficiency.
- Qwen Sparse Attention: QSA is one of the two architectural changes publicly highlighted for Flash-Next. Its detailed implementation and quantitative benefits remain to be documented by Qwen.
- Open-weight ecosystem orientation: The release is intended to give the open-source community early access to the architectural changes so inference runtimes and downstream tooling can adapt before Qwen4 arrives.
- Coding and Agent Performance of Qwen3.8-Flash-Next
| Benchmark | Qwen3.8-Flash-Next | Qwen3.8-27B | Qwen3.7-Plus | DeepSeek V4 Flash | Claude Opus 4.6 |
|---|---|---|---|---|---|
| DeepSWE 1.1 | 58.7 | 42.2 | 16.5 | 54.4 | -- |
| SWE-bench Pro | 62.5 | 61.7 | 55.8 | 56.0 | 53.4 |
| SWE-bench Multilingual | 81.0 | 73.8 | 75.8 | -- | 77.5 |
| NL2Repo-Bench | 48.1 | 42.3 | 41.1 | 54.2 | 47.6 |
| CoWorkBench | 73.9 | 70.7 | 65.1 | 45.1 | 68.2 |
| JobBench | 55.7 | 33.4 | 27.6 | 41.3 | 36.6 |
| Toolathlon Verified | 73.5 | 67.1 | 50.6 | 70.3 | -- |
| IFBench | 81.3 | 79.5 | 79.1 | 79.2 | 62.5 |
| 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 |
Qwen3.8-Flash-Next vs Qwen3.8-Max vs Qwen3.8-27B
| Model | Positioning | Architecture / Scale | Multimodal | Current status |
|---|---|---|---|---|
| Qwen3.8-Flash-Next | Qwen4 architecture preview / efficiency-oriented open model | Multimodal MoE; complete specifications pending | Yes | Upcoming |
| Qwen3.8-Max | Flagship hosted Qwen3.8 model | Large-scale MoE | Yes | Available |
| Qwen3.8-27B | Open-weight Qwen3.8 model | 27B dense model | Model-specific capabilities | Available |
Qwen3.8-Max is already available through Alibaba's cloud ecosystem and is positioned for advanced reasoning, visual understanding, coding, and agentic workloads. Qwen's current documentation lists Qwen3.8-Max with a 1M-token context window, while CometAPI already exposes qwen3.8-max.
Flash-Next should therefore be viewed differently: its primary significance at this stage is architectural rather than benchmark-based. It previews the technical direction of Qwen4 and gives the open-source ecosystem an earlier target for runtime and deployment optimization.