Qwen 3.8 Flash

Alibaba Qwen open-sourced Qwen3.8-Flash as a multimodal MoE with 125B parameters plus 51B n-gram embeddings and only 6B active per token, alongside an early-look Qwen3.8-Flash-Next variant previewing the Qwen4 architecture. The report cites benchmark scores of 58.7 on DeepSWE 1.1, 62.5 on SWE-bench Pro, and 73.9 on CoW The architecture introduces a GDN+QSA hybrid attention layer with Gated Residual, N-gram embeddings, and a Muon optimizer, where QSA delivers up to 7.6x faster prefill at 1M-token context. Weights were released on Hugging Face and ModelScope with day-0 support in vLLM and SGLang, and the QwenCloud API is priced at $0.1

A Reuters-sourced report via The Economic Times confirms Alibaba's Qwen team released Qwen3.8-Flash as a multimodal model serving stronger coding and office-task performance while cutting training costs relative to the prior Qwen3.7-Plus. The piece explicitly names the subject model Qwen3.8-Flash and reports a default The same report notes Alibaba simultaneously released open-source weights for Qwen3.8-Flash-Next, framed as an architecture prototype for its next-generation Qwen4 model family, giving developers an early look at the Qwen4 design. The release comes amid intensifying domestic AI competition and follows Alibaba's heavily

Vercel announced in its official changelog on August 26, 2026 that Qwen 3.8 Flash from Alibaba is now available on AI Gateway. The model accepts text and images as input, offers a 1 million token context window, and can return up to 65k tokens per response. Alibaba positions it for coding, tool use, and multi-step agen To use the model, developers set `model` to `alibaba/qwen3.8-flash` in the AI SDK, or pick it inside their coding agent after running `vercel ai-gateway coding-agents setup`. A playground is available for testing. Vercel notes AI Gateway provides a unified API with retries, failover, performance optimizations, Zero Dat

A third-party technical write-up describes Qwen3.8-Flash and the simultaneously open-sourced Qwen3.8-Flash-Next, an early preview of the Qwen4 architecture. The base model is a multimodal MoE with 125B total parameters plus 51B N-gram Embedding and 6B parameters activated per token, natively supporting 262K tokens and Four architectural upgrades are detailed: a GDN+QSA hybrid attention module that achieves up to 7.6x and 4.9x speedups in Prefill and Decode respectively at million-token lengths; a Gated Residual mechanism expanding the residual flow into four parallel branches; N-gram Embedding via local context lookup tables adding

Alibaba officially unveiled Qwen3.8-Flash as an open-weight, multimodal Mixture-of-Experts model positioned as an early preview of the architecture that will power the upcoming Qwen4 series. The model features a 125B-parameter main model supplemented by 51B N-gram embeddings, with only 6B parameters activated per token On benchmarks, Alibaba reports Qwen3.8-Flash performs competitively against leading models such as DeepSeek-V4-Flash and Claude-Opus-4.6 across SWE-bench Pro, CoWorkBench, Toolathlon Verified, MathVision, AndroidWorld, and ERQA, demonstrating strengths in agentic coding, long-horizon agent tasks, and multimodal intelli

TypingMind's changelog on August 27, 2026 lists Qwen 3.8 Flash as a newly added AI model across its supported providers, alongside GLM-5.3 Flash (Ox Alpha). The entry introduces Qwen 3.8 Flash as the latest model from Alibaba, signaling third-party ecosystem adoption within a day of Vercel's announcement. TypingMind's same changelog also documents related fixes and improvements, including native PDF capability detection for custom models with a new "Support PDF Input" toggle, a fix for synced images not working with the DeepSeek provider, and a fix for attachments not being accessible in forked chats. These additions ar

The Dev Signal newsletter "Gateway Rush" covers the Qwen 3.8 Flash launch on Vercel AI Gateway as one of three gateway additions. It characterizes the model as a 1M-token context, 65k output window multimodal model integrated into Vercel's unified inference API with no platform fee, framing it as a cost-competitive, lo The newsletter also points developers to existing AI Gateway tooling for unified API calls, usage and cost tracking, retries, and failover, and notes the same pricing passthrough with no markup. As an aggregator, the piece adds editorial framing rather than new primary facts, but reinforces that Qwen 3.8 Flash is posit