Qwen3.8 Flash

Alibaba released Qwen3.8-Flash on August 26, 2026, making it downloadable and publishing its weights. The 125B-parameter model uses an architecture designed for the next-generation Qwen 4 series, significantly reducing training and inference costs. Alibaba positions it as a lower-priced option to drive global adoption of its AI lineup. The release is competitive with Anthropic's Opus 4.6 and DeepSeek's V4-Flash, according to Alibaba. The company has pledged more than 380 billion yuan (about US$57 billion) over three years toward AI development and infrastructure, funding the effort in part with a US$10.2 billion Hong Kong follow-on share offering announced the same week.

Qwen3.8-Flash is a 125B-parameter multimodal MoE with 51B n-gram embeddings and 6B active parameters per token, trained at roughly one-ninth the cost of Qwen3.7-Plus. It ships as an early look at the Qwen4 architecture, featuring GDN+QSA hybrid attention, Gated Residual, n-gram embeddings, and Muon training, with QSA delivering up to 7.6x faster 1M-token prefill. The model scores 58.7 on DeepSWE 1.1, 62.5 on SWE-bench Pro, and 73.9 on CoWorkBench, beating Qwen3.7-Plus on coding and office work benchmarks. Weights are available on Hugging Face and ModelScope with day-0 vLLM and SGLang support, and QwenCloud API pricing is $0.16 per million input tokens and $0.47 per million output tokens.

The LLM-Stats scorecard for Qwen3.8 Flash, dated August 26, 2026, places the model at $0.17 per million blended tokens with an LLM Stats Score of 49.1, positioning it between DeepSeek-V4-Flash-0731 ($0.066, score 44.7) and GPT-6 Astra ($11.9, score 59.9) on the cost-efficiency chart. Quality Tracker signals show the mo Performance-by-conversation-depth visualization on the same page tracks how Qwen3.8 Flash holds up as conversations get longer, with the chart spanning September 1 through September 14, 2026, ranging between -6.0σ and +6.0σ and showing a positive trajectory over the period. The scorecard explicitly identifies the subje

Alibaba released Qwen3.8-Flash on August 26, 2026, as an open-weight, multimodal Mixture-of-Experts model with 125B parameters plus 51B N-gram embeddings and 6B parameters activated per token. The model serves as an early architecture preview of the upcoming Qwen4 series, targeting high-volume applications, tool-driven workflows, and coding or co-working assistants. Qwen3.8-Flash natively supports 262K tokens of context extensible to 1M, performs competitively against DeepSeek-V4-Flash and Claude-Opus-4.6 on benchmarks including SWE-bench Pro, CoWorkBench, Toolathlon Verified, MathVision, AndroidWorld, and ERQA, and ships weights on Hugging Face and ModelScope with API access at 1 RMB input / 3 RMB output per 1M tokens on Model Studio and Qwen Cloud.