Qwen3.8-Flash-Next
A lightweight inference model in Alibaba's Qwen series, supporting switchable thinking and non-thinking modes, optimized for fast inference scenarios.
Core Facts
Qwen3.8-Flash-Next adopts a Mixture of Experts (MoE) architecture with 125B total parameters and 6B activated parameters.
In an RTX 5070 (12GB VRAM) configuration, the local Qwen3 9B model's output speed reached approximately 70–76 tokens/second.
A unified memory pool of 128GB means that quantized models at the 70-billion-parameter scale or larger can be run locally.
Timeline (last 90 days)
Qwen3.8-Flash-Next on MLX is up to 1.4 times faster on M5 Macs and up to 1.2 times faster on M1 through M4.
The original Qwen3.8-Flash-Next weights support vLLM, SGLang, Transformers and TokenSpeed, and SGLang had day-0 support on the day of release.
No vision benchmark at all was released for the vision part of NVIDIA's FP4 version of Qwen3.8, and the distillation data was all text.
NVIDIA's FP4 version of Qwen3.8 achieves 91.5% accuracy on GPQA Diamond, versus 92.2% for the BF16 baseline, a difference of 0.7 percentage points.
The Qwen3.8-Flash-Next FP4 version released by NVIDIA is about 96 GiB in size, while the original BF16 is 355 GB.
The Qwen3.8-Flash-Next model supports an ultra-long context window of 262K tokens.
Qwen3.8-Flash-Next uses the Strata inference engine in conjunction with MTP (Multi-Token Prediction) acceleration technology.
A 262K token context means the model can process approximately 200,000 Chinese characters in a single conversation, at the cost of the KV cache occupying large amounts of VRAM and memory as the context grows.
The Flash and Next suffixes in the model name are not standard naming conventions from the official Qwen version series.
The model successfully fixed two bugs in a Snake HTML game (the snake not dying upon hitting a wall and the snake being unable to eat food).
40 more timeline events
All Facts (20)
Qwen3.8-Flash-Next adopts a Mixture of Experts (MoE) architecture with 125B total parameters and 6B activated parameters.
85%VerifiedIn an RTX 5070 (12GB VRAM) configuration, the local Qwen3 9B model's output speed reached approximately 70–76 tokens/second.
70%VerifiedA unified memory pool of 128GB means that quantized models at the 70-billion-parameter scale or larger can be run locally.
70%VerifiedMemory requirements are at least 37.6 GB for Q2_0, at least 39.2 GB for IQ2_XS, and at least 47 GB for IQ3_XXS, with the vision encoder occupying an additional 0.91 GB.
65%VerifiedQwen3.8-Flash-Next achieved scores of 58.7 and 62.5 on the SWE-bench 1.1 and SWE-bench Pro evaluations, respectively.
65%VerifiedBoth the project code and model weights have been open-sourced.
75%UnverifiedThrough MTP (Multi-Token Prediction) technology, the inference speed of Qwen3.8-Flash-Next and GLM-5.3-Flash is improved by up to 2x.
70%UnverifiedQwen3.8-Flash-Next has a total parameter count of 125B with approximately 6B activated parameters, and belongs to the sparse MoE architecture.
60%UnverifiedQwen3.8-Flash-Next has approximately 6B activated parameters.
60%UnverifiedIf the quantized model weights are 150 GB, reading the weights once at a bandwidth of 273 GB/s takes approximately 0.55 seconds, giving a theoretical single-token generation speed of approximately 1.8 tokens/s (excluding computation latency and cross-node communication overhead).
60%UnverifiedQwen3.8-Flash-Next on MLX is up to 1.4 times faster on M5 Macs and up to 1.2 times faster on M1 through M4.
50%UnverifiedThe Qwen3.8-Flash-Next FP4 version released by NVIDIA is about 96 GiB in size, while the original BF16 is 355 GB.
50%UnverifiedNVIDIA's FP4 version of Qwen3.8 achieves 91.5% accuracy on GPQA Diamond, versus 92.2% for the BF16 baseline, a difference of 0.7 percentage points.
50%UnverifiedNo vision benchmark at all was released for the vision part of NVIDIA's FP4 version of Qwen3.8, and the distillation data was all text.
50%UnverifiedThe original Qwen3.8-Flash-Next weights support vLLM, SGLang, Transformers and TokenSpeed, and SGLang had day-0 support on the day of release.
50%UnverifiedThe Flash and Next suffixes in the model name are not standard naming conventions from the official Qwen version series.
50%UnverifiedA 262K token context means the model can process approximately 200,000 Chinese characters in a single conversation, at the cost of the KV cache occupying large amounts of VRAM and memory as the context grows.
50%UnverifiedThe Qwen3.8-Flash-Next model supports an ultra-long context window of 262K tokens.
50%UnverifiedQwen3.8-Flash-Next uses the Strata inference engine in conjunction with MTP (Multi-Token Prediction) acceleration technology.
50%UnverifiedQwen3.8-Flash-Next has a 48-layer structure with 512 experts per layer and 180B-class parameters (125B main body + 51B N-gram table + 4B MTP).
50%