Qwen Flash Next
A lightweight, efficiency-oriented model in Alibaba's Qwen series, designed to achieve high inference performance at a relatively small parameter scale. It supports deployment in multiple quantization formats.
Timeline (last 90 days)
A post appeared on Reddit claiming to have conclusive evidence that Qwen Flash Next tried to send data to a remote Chinese server, but it was actually a system prompt phishing prank.
The official team provides corresponding GGUF quantized model files on Hugging Face, with the repository address at ggml-org/Qwen3.8-Flash-Next-GGUF.
Specific performance improvement figures, compatible hardware ranges, and actual MTP acceleration ratios still require confirmation from official documentation and subsequent community benchmarking.
Work on MTP (multi-token prediction) support for the Qwen Flash Next model has been restarted.
Developers have prepared switchable quantized models for Qwen Flash Next and submitted a corresponding Pull Request to the llama.cpp repository.
The MTP support work is currently still in a WIP (work in progress) state and has not yet stabilized.
The quantized model for Qwen Flash Next is hosted on Hugging Face in the ggml-org/Qwen3.8-Flash-Next-GGUF repository, providing weight files in GGUF format.
To complete the 250-case test, FP8 27B took 36 minutes and 27 seconds, while Flash-Next IQ4_XS took 2 hours, 5 minutes, and 47 seconds, a difference of approximately 3.5 times.
On GSM8K, Flash-Next scored 97.5% versus FP8's 90.0%; on German MGSM, 92.5% versus 90.0%; and on IFEval, both tied at 89.6%.
On the MMLU-Pro benchmark, Flash-Next IQ4_XS scored 83.8% and Qwen3 27B FP8 scored 75.0%.
3 more timeline events
All Facts (6)
Specific performance improvement figures, compatible hardware ranges, and actual MTP acceleration ratios still require confirmation from official documentation and subsequent community benchmarking.
50%UnverifiedThe tester recommends choosing the lower-quantized Flash Next for higher accuracy in single-user scenarios, and choosing FP8 27B for concurrent processing since its speed is approximately 4 times that of QFN.
50%UnverifiedAMD's RX 9700 series GPUs have hardware-level acceleration support for mxfp4, and upgrading to 128GB of memory and using the mxfp4 quantized version of Flash Next can increase QFN speed by approximately 2 times.
50%UnverifiedA post appeared on Reddit claiming to have conclusive evidence that Qwen Flash Next tried to send data to a remote Chinese server, but it was actually a system prompt phishing prank.
50%UnverifiedThe official team provides corresponding GGUF quantized model files on Hugging Face, with the repository address at ggml-org/Qwen3.8-Flash-Next-GGUF.
50%UnverifiedThe quantized model for Qwen Flash Next is hosted on Hugging Face in the ggml-org/Qwen3.8-Flash-Next-GGUF repository, providing weight files in GGUF format.
50%