Qwen3.8 Flash Next: Open Weights and Multimodal Preview

Release Details and Architectural Context

The official announcement occurred on August 26, 2026. This date marks the historical event of the release, distinct from the current publication date of this article. Qwen identified Qwen3.8 Flash Next as a multimodal Mixture-of-Experts (MoE) model. The provider confirmed that the weights are open, allowing access to the model parameters.

The source states that this model serves as an early preview of the architecture used in Qwen4. It plays a similar role to Qwen3-Next within the Qwen3.5 lineage. The documentation notes that the hybrid Gated DeltaNet + Gated Attention design introduced during that earlier period has been utilized across the Qwen3.5, Qwen3.6, Qwen3.7, and Qwen3.8 series.

Input Modalities and Output Capabilities

According to the verified facts, the model accepts both text and image inputs. Its output modality is restricted to text. No specific context window size or maximum output token limit is provided in the source material. Pricing information is also absent from the available data.

The model card confirms the existence of the repository under the Qwen organization. Multiple commits are referenced in the evidence, indicating active development or versioning on the platform. However, no performance benchmarks, latency metrics, or accuracy scores are included in the source facts.

Evaluation Procedure for Engineers

Engineers can evaluate the model by conducting controlled tests using the open weights. First, test the model’s ability to process mixed text and image inputs against a defined dataset. Measure the inference time and resource consumption during these runs. Compare the generated text outputs against ground truth data to validate accuracy. Finally, assess whether the hybrid attention mechanism behaves consistently across different input types.

It is important to distinguish between documented features and unverified benefits. The source confirms the architectural components and modalities but does not claim specific improvements over previous models. Therefore, any evaluation should focus on empirical results rather than assumed advantages.

• Test multimodal input handling with standardized datasets.
• Measure inference latency and memory usage.
• Compare output quality against baseline models.
• Validate consistency of the hybrid attention design.

Artificial Analysis benchmark review

Benchmark review updated on 2026-10-12. These scores apply to the linked artifact and recorded methodology; they do not establish the performance of a different runtime or quantized model.

Review the full benchmark provenance and source methodology →

Primary sources

• Qwen original announcement ↗

Continue your research

• Qwen3.8 Flash Next model evidence →
• Compare Qwen3.8 Flash Next with another model →
• Current model comparisons →
• How to compare models for a workflow →
• Get source-backed release updates →