Qwen3.8 Flash Next
A lightweight model variant within Alibaba's Qwen3 series, designed for efficient inference scenarios. It supports community quantization and local deployment.
Core Facts
The approximately 20:1 ratio between Qwen3 Next Flash's 125 billion total parameters and 6 billion activated parameters represents a fairly aggressive sparsity design.
In an RTX 5070 (12GB VRAM) configuration, the local Qwen3 9B model's output speed reached approximately 70–76 tokens/second.
The benchmark results of Qwen3.8 Flash Next are official Qwen data and have yet to be independently verified by third parties.
Timeline (last 90 days)
The main model of the architecture used by Qwen4 has 125B parameters, activates only 6B per token, uses a MoE architecture, contains 512 experts, and has a native context of 262K.
The license for Qwen3.8 Flash Next is the Qwen Community Model License rather than Apache 2.0, and the terms need to be confirmed before commercial use.
When Qwen3.8 Flash Next is deployed locally, latency increases by only 6% with 64G of memory; when compressed to 24G it uses only 6G of VRAM, with latency steady at 9.1 milliseconds.
The Ngram table of Qwen3.8 Flash Next is a separate 47.7G file, larger than the 10G that is activated.
Qwen3.8 Flash Next has reached 3.63 million downloads on Hugging Face.
Tested on a machine with an AMD 9700 + 32GB of VRAM + 64GB of memory, the model can run a 56K context.
The quantized Coder model deployed with this approach has limited text generation quality and is suitable for low-cost experimentation or as a free token source for agents.
According to benchmark data from Hugging Face, the IQ3_S tier scores an average of 93.26, close to the full-precision base model's score of 93.12.
Qwen3.8 Flash Next has 125B total parameters, activates only approximately 6B per token, supports a native context of 260,000 tokens expandable to 1 million, and its training cost is reportedly only one-ninth that of the previous generation.
Qianwen officially placed Qwen3.8 Flash Next into a subscription plan to sell API access externally, while the 27B version has not yet been sold separately.
28 more timeline events
All Facts (20)
The approximately 20:1 ratio between Qwen3 Next Flash's 125 billion total parameters and 6 billion activated parameters represents a fairly aggressive sparsity design.
70%VerifiedIn an RTX 5070 (12GB VRAM) configuration, the local Qwen3 9B model's output speed reached approximately 70–76 tokens/second.
70%VerifiedThe benchmark results of Qwen3.8 Flash Next are official Qwen data and have yet to be independently verified by third parties.
65%UnverifiedQwen3.8 Flash Next is equipped with an independent Ngram embedding system that adds approximately 51 billion additional parameters; these parameters do not participate in regular forward propagation and are activated only upon specific pattern matches
60%UnverifiedQwen3.8 Flash Next supports two structured output modes: function calling and constrained generation.
60%UnverifiedThe license for Qwen3.8 Flash Next is the Qwen Community Model License rather than Apache 2.0, and the terms need to be confirmed before commercial use.
50%UnverifiedThe main model of the architecture used by Qwen4 has 125B parameters, activates only 6B per token, uses a MoE architecture, contains 512 experts, and has a native context of 262K.
50%UnverifiedQwen3.8 Flash Next has reached 3.63 million downloads on Hugging Face.
50%UnverifiedThe Ngram table of Qwen3.8 Flash Next is a separate 47.7G file, larger than the 10G that is activated.
50%UnverifiedWhen Qwen3.8 Flash Next is deployed locally, latency increases by only 6% with 64G of memory; when compressed to 24G it uses only 6G of VRAM, with latency steady at 9.1 milliseconds.
50%UnverifiedTested on a machine with an AMD 9700 + 32GB of VRAM + 64GB of memory, the model can run a 56K context.
50%UnverifiedThe quantized Coder model deployed with this approach has limited text generation quality and is suitable for low-cost experimentation or as a free token source for agents.
50%UnverifiedAccording to benchmark data from Hugging Face, the IQ3_S tier scores an average of 93.26, close to the full-precision base model's score of 93.12.
50%UnverifiedQwen3.8 Flash Next has 125B total parameters, activates only approximately 6B per token, supports a native context of 260,000 tokens expandable to 1 million, and its training cost is reportedly only one-ninth that of the previous generation.
50%UnverifiedStrata enables a single GPU with 12 GB of VRAM to smoothly run Qwen3.8 Flash Next, a large model with 125 billion activated parameters and an overall size of 176B.
50%UnverifiedWhen extending the context from 192,000 to 256,000, the expert cache dropped from 16.44 GB to 13.66 GB and KV increased from 4.31 GB to 7 GB, with almost zero speed cost, because KV is fully in VRAM and the expert cache automatically balances.
50%UnverifiedOn a single RX 7900 XTX 24G GPU, the Strata engine can fully run the approximately 125B-parameter Qwen3.8 Flash Next MoE model, supporting both text inference and image recognition.
50%UnverifiedActual testing of two quantization models: IQ2_XXS achieves 48 tokens/s for short Q&A and 63 tokens/s for long text; IQ3_XXS achieves 30 tokens/s for short Q&A and 39 tokens/s for long text, with VRAM usage of approximately 16.44 GB, activating 10,138 experts—IQ2 is about 60% faster than IQ3.
50%UnverifiedStrata is an open-source inference runtime capable of running the 125-billion-parameter Qwen3.8 Flash Next model on an RTX 5070 (12 GB VRAM) at an output speed of approximately 90 tokens per second.
50%UnverifiedIn Qwen3.8 Flash Next, each layer has a router that selects 10 active experts out of 512, plus one shared expert that always runs.
50%