Qwen3.8 Flash
A multimodal language model in Alibaba Cloud's Qwen series, employing a sparse Mixture-of-Experts (MoE) architecture with 125 billion total parameters of which approximately 6 billion are activated per inference. It natively supports text, image, and video inputs as well as million-token context windows.
Core Facts
Qwen 3.8 Flash Next adopts a Mixture of Experts (MoE) architecture.
Strata uses low-bit quantization; the minimum configuration requires approximately 38 GB of combined memory and VRAM, the highest-quality configuration requires approximately 55 GB, and 64 GB of system memory is the ideal target.
Developer Primitive reported a generation speed of 142.6 tokens/s for Qwen3 Flash Next running locally.
Timeline (last 90 days)
A comparative test shows that the Anslows UD-Q4 quantized version of Qwen3.8 Flash (125B) scores higher overall than the official Flash API in engineering delivery scenarios.
In a Python programming problem test, Strata completed it in approximately 4 seconds while the cloud took nearly 19 seconds; in a creative writing test, Strata took approximately 3 seconds and the cloud approximately 10 seconds
Qwen3.8 Flash is a MoE model with 24,576 experts across 48 layers internally, activating only approximately 10 experts per generated token.
Strata uses low-bit quantization; the minimum configuration requires approximately 38 GB of combined memory and VRAM, the highest-quality configuration requires approximately 55 GB, and 64 GB of system memory is the ideal target.
Developer Primitive reported a generation speed of 142.6 tokens/s for Qwen3 Flash Next running locally.
On a 128G compute module device, the decoding speed of Qwen3.8 Flash routinely reaches 241 TPS, with a peak of 274 TPS.
Qwen3.8 Flash supports a thinking mode (reasoning before answering) and can also disable thinking to answer directly.
The author proposes three metrics for evaluating multimodal long-context models: whether key information in long materials is missed, whether numbers in images are misread, and whether the output format is directly usable.
Qwen3.8 Flash is a native multimodal model that supports text, images, and video as inputs and outputs text.
Qwen3.8 Flash scored 73.9 on Cowork Bench, while DeepSeek scored 45.1.
16 more timeline events
All Facts (20)
Qwen 3.8 Flash Next adopts a Mixture of Experts (MoE) architecture.
80%VerifiedStrata uses low-bit quantization; the minimum configuration requires approximately 38 GB of combined memory and VRAM, the highest-quality configuration requires approximately 55 GB, and 64 GB of system memory is the ideal target.
70%VerifiedDeveloper Primitive reported a generation speed of 142.6 tokens/s for Qwen3 Flash Next running locally.
65%VerifiedThe cloud production version Qwen3.8 Flash provides a 1 million token context by default and has the official toolchain built in.
65%UnverifiedQwen3.8 Flash is a native multimodal model that supports text, images, and video as inputs and outputs text.
60%UnverifiedA comparative test shows that the Anslows UD-Q4 quantized version of Qwen3.8 Flash (125B) scores higher overall than the official Flash API in engineering delivery scenarios.
50%UnverifiedIn a Python programming problem test, Strata completed it in approximately 4 seconds while the cloud took nearly 19 seconds; in a creative writing test, Strata took approximately 3 seconds and the cloud approximately 10 seconds
50%UnverifiedQwen3.8 Flash is a MoE model with 24,576 experts across 48 layers internally, activating only approximately 10 experts per generated token.
50%UnverifiedOn a 128G compute module device, the decoding speed of Qwen3.8 Flash routinely reaches 241 TPS, with a peak of 274 TPS.
50%UnverifiedQwen3.8 Flash supports a thinking mode (reasoning before answering) and can also disable thinking to answer directly.
50%UnverifiedThe author proposes three metrics for evaluating multimodal long-context models: whether key information in long materials is missed, whether numbers in images are misread, and whether the output format is directly usable.
50%UnverifiedThe activated parameter count of Qwen3.8 Flash is approximately one-third that of GLM.
50%UnverifiedThe cloud production version of Qwen3.8 Flash provides a default context of 1 million tokens.
50%UnverifiedQwen3.8 Flash excels in office-type collaborative scenarios, while GLM excels in programming Agent scenarios.
50%UnverifiedIn production environments, max_tokens limits should be set, concise output should be requested in the system prompt, and structured output formats should be used to constrain responses in order to control costs.
50%UnverifiedAlibaba's official figures show Qwen3.8 Flash scoring 62.5 on SWE-Bench Pro, leading Claude Opus 4.6 by approximately 9 points; 73.9 on Cowork Bench versus DeepSeek's 45.1; and 55.7 on Job Bench, exceeding Opus 4.6 by approximately 20 points.
50%UnverifiedThose pursuing the ultimate coding agent experience or the fastest possible response speed should think carefully before choosing Qwen3.8 Flash and may consider GLM or DeepSeek instead.
50%UnverifiedQwen3.8 Flash is divided into two products: the open-weight version Qwen3.8 Flash Next and the cloud production version Qwen3.8 Flash.
50%UnverifiedQianwen's output volume is significantly higher than the platform median, and verbose output drives up total billing costs.
50%UnverifiedModel capability, generation speed, and agent framework scheduling are three non-equivalent dimensions; strong model capability does not mean fast perceived responsiveness.
50%