Open weights land as a dry run for Qwen4

Alibaba has released Qwen3.8-Flash, a 125-billion-parameter open-weight multimodal Mixture-of-Experts model the company presents as an early preview of the architecture behind Qwen4. The model follows Qwen 3.8 Max, which arrived at the start of the month, and was detailed in a release blog covered by The New Stack.

At a glance

• Alibaba says the release plays the same role Qwen3-Next played for Qwen3.5, when a hybrid Gated DeltaNet and Gated Attention design was introduced ahead of the Qwen3.5 through Qwen3.8 series.
• Access runs through Alibaba's Model Studio and Qwen Cloud at $0.16 per million input tokens and $0.47 per million output tokens, with weights published on Hugging Face and ModelScope.
• Alibaba claims flagship-level capability at everyday cost, with a redesigned standard mode in QwenWork cutting token consumption per task by 75% and roughly doubling generation speed compared with the current mode.

Releasing the architecture ahead of the model family reads as an attempt to have the community stress-test Gated Residual, N-gram embeddings and sparse attention before Qwen4 is built on top of them. It also gives teams something concrete to plan codebases against. The cost claim likely matters more than the benchmark margins: a model trained on a fraction of the resources of a predecessor three times its size shifts what smaller teams can attempt.

Qwen3.8-Flash-Next activates 6 billion of its 125 billion parameters per token

Qwen3.8-Flash-Next is the open-weight research-frontier model, published on Hugging Face and Alibaba's ModelScope community portal. Qwen3.8-Flash, built on the same underlying architecture, is the production version, served to developers through the QwenCloud API with 1 million tokens by default and official built-in tools.

The open model pairs a 125-billion-parameter main network with an additional 51 billion N-gram embedding parameters, and activates 6 billion parameters per token. Native context is 262,144 tokens, extendable to 1,000,000 tokens with YaRN. Alibaba positions the model as an optimal balance among capability, latency and cost, claiming superior capabilities in coding and office tasks.

In this release, we are opening the weights of Qwen3.8-Flash-Next, a multimodal MoE model that also serves as an early preview of the architecture used in Qwen4.

The company says the release plays the same role Qwen3-Next played for Qwen3.5, when the hybrid Gated DeltaNet and Gated Attention design first reached developers. That design then ran through the Qwen3.5, Qwen3.6, Qwen3.7 and Qwen3.8 series, and the new model is framed as the equivalent groundwork for Qwen4.

Alibaba upgraded the model along four axes: attention, residual, embedding and optimization

The hybrid attention design combines Gated DeltaNet, which compresses historical information, with Qwen Sparse Attention, which uses a lightweight compressed indexer to pick out relevant context. Alibaba says the pairing substantially reduces attention costs on long sequences. The company describes Qwen Sparse Attention as a novel design.

A Gated Residual mechanism expands the data pathways between layers, which Alibaba says strengthens cross-layer information flow and training stability. N-gram Embedding scales model capacity with minimal additional computation, and the Muon optimizer improves the efficiency of large-scale model training, according to the company.

Alibaba says the model needs around one-ninth of the training resources while outperforming Qwen3.7-Plus, a model three times its size, and that this significantly reduces both training and inference costs. The company also states that developers can run the model on hardware they are likely to own already.

Qwen3.8-Flash-Next scores 62.5 on SWE-bench Pro, ahead of Claude-Opus-4.6

Alibaba benchmarked the model on SWE-bench Pro for agentic coding, CoWorkBench for long-horizon office work, Toolathlon Verified for real-world tool use, MathVision for visual maths problem-solving, AndroidWorld for agentic mobile use and ERQA for embodied intelligence, against models including DeepSeek-V4-Flash and Claude-Opus-4.6.

On SWE-bench Pro, Qwen3.8-Flash-Next scores 62.5, against 61.7 for Qwen3.8-27B, 56.0 for DeepSeek-V4-Flash-0731, 55 for Qwen3.7-Plus and 53.4 for Claude-Opus-4.6 (Max). Mechatronics engineer Alok wrote on X that he ran the 125B A6B model with a 250,000 token context window on a single 24GB RTX 4090.

Alok reported 21 tokens per second decode and 364 tokens per second prefill without MTP or KV cache quantization. Developer reaction has been split: a software developer posting on Hacker News as Embedding Shapes said the sentiment around Qwen 3.8 raised hopes, but local models are not yet adequate for professional software development.

No date fixed for Qwen4

Alibaba has not confirmed a launch date for Qwen4. Commentators and market watchers broadly expect it before the end of the year, possibly as early as September, though nothing has been announced. Speculation among those watchers centres on wider Mixture-of-Experts scaling, deeper native multimodal processing and optimization for complex 3D coding and advanced spatial modelling tasks.