Qwen3.8-Flash open-sourced, surpassing V4 Flash: early adoption of next-generation Qwen4 architecture

Tonight, in addition to Zhipu open-sourcing the Niu Lai large model, Alibaba's Qwen also released the Qwen3.8-Flash large model. Although it still uses the 3.X version name, it is actually based on the next-generation Qwen4 architecture.

Qwen previously explained that Qwen3.8-Flash is designed to let the open-source community get an early experience of the new technologies in the Qwen4 architecture, serving as a technical preview of the next-generation model.

According to official information, Qwen3.8-Flash is a multimodal MoE model with 125B parameters plus 51B N-gram embeddings, activating only 6B per token, with a 262K native context that can be extended to 1M via YaRN.

The company states it offers unmatched cost-effectiveness, priced at just $0.16 per 1M input tokens and $0.47 per 1M output tokens.

For comparison, DS V4 Flash currently costs $0.22 and $0.66 respectively, and prices double during peak hours, so Qwen3.8-Flash is indeed highly cost-effective.

Technically, Qwen3.8-Flash adopts the next-generation architecture—GDN + QSA hybrid attention, gated residual connections, N-gram embeddings, and the Muon optimizer—serving as the precursor to the architecture used in Qwen4.

In terms of performance benchmarks, Qwen3.8-Flash scores 58.7 on DeepSWE 1.1, 62.5 on SWE-bench Pro, 73.9 on CoWorkBench, 84.5 on AndroidWorld, and 95.7 on MathVision (with CI).

Overall, this performance surpasses V4 Flash, and with its high cost-effectiveness, if real-world performance matches these numbers, it might be worth switching to this large model.