'Qwen3.8-Flash-Next' is released as an open model, using the next-generation architecture of Qwen4.

Alibaba's AI research team, Qwen, released its AI model ' Qwen3.8-Flash-Next ' on August 26, 2026. Qwen3.8-Flash-Next was developed using the next-generation architecture planned for use in the Qwen4 series, and has succeeded in building a high-performance model while keeping training costs down. A quantized version of the model using Unsloth has also already been released.

Qwen3.8-Flash-Next: A New Architecture, Towards Ultimate Cost-Efficiency

⚡Meet Qwen3.8-Flash, a multimodal MoE and an early preview of the Qwen4 architecture, now open-weight!

The production version Qwen3.8-Flash will be available soon via QwenCloud API at just $ 0.16/1M input tokens and $ 0.47/1M output tokens.

125B parameters + 51B N-gram… pic.twitter.com/SScnmzWS7O

— Qwen (@Alibaba_Qwen) August 26, 2026

Qwen3.8-Flash-Next employs a hybrid architecture incorporating 'Gated DeltaNet (GDN)' for efficient history compression and 'Qwen Sparse Attention (QSA)' for significantly reducing the cost of attention processing. It also incorporates 'N-gram Embedding,' which allows the model's representable vocabulary to be expanded with minimal additional computation. The total number of parameters is 125 billion, the number of active parameters is 6 billion, and the number of N-gram Embedding parameters is 51 billion. It supports a context length of 262,144 tokens by default, and can support up to 1 million tokens with 'YaRN,' which extends the context length.

The benchmark results for 'Qwen3.8-Flash-Next', 'Qwen3.8-27B', 'Qwen3.7-Plus', 'DeepSeek-V4-Flash-0731', and 'Claude-Opus-4.8(Max)' are shown below. Qwen3.8-Flash-Next outperforms Claude-Opus-4.8(Max) in many tests. Furthermore, the training period required for Qwen3.8-Flash-Next is said to be one-ninth that of Qwen3.8-27B.

Qwen3.8-Flash-Next is also characterized by its ability to minimize throughput degradation in both prefill and decode when processing large tokens. The graph below shows the difference in prefill speed for different context lengths, with Qwen3.7-Plus as the baseline. When processing 1 million tokens, Qwen3.8-Flash-Next is 8.6 times faster than Qwen3.7-Plus.

Qwen3.8-Flash-Next is available as an open model on Hugging Face and ModelScope.

Qwen/Qwen3.8-Flash-Next · Hugging Face
https://huggingface.co/Qwen/Qwen3.8-Flash-Next

千问3.8-Flash-Next · Model Tower
https://modelscope.cn/models/Qwen/Qwen3.8-Flash-Next

Additionally, a quantized version by Unsloth was released on the same day as the announcement. The 1-bit version can run on machines with 75GB or more of RAM and maintains 80% of the accuracy of the full model. Unsloth's documentation boasts that 'even when loaded into RAM and executed, it can achieve the same performance as when loaded into VRAM.'

unsloth/Qwen3.8-Flash-Next-GGUF · Hugging Face
https://huggingface.co/unsloth/Qwen3.8-Flash-Next-GGUF