Qwen3 Next Flash
A large-scale Mixture-of-Experts (MoE) language model released by Alibaba, with a total parameter count at the 125B scale. It uses a sparse activation architecture to reduce per-inference computational cost and supports efficient inference on consumer-grade hardware.
Core Facts
The 4-bit quantized version (Q4K XL) of Qwen3 Next Flash is approximately 111 GB.
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.
Timeline (last 90 days)
The Ngram embedding table of Qwen3 Next Flash can be offloaded to host memory (CPU RAM) and executed with overlap with model computation via asynchronous prefetching.
The 4-bit quantized version (Q4K XL) of Qwen3 Next Flash is approximately 111 GB.
Qwen3 Next Flash's API pricing is $0.16 per million input tokens and $0.47 per million output tokens.
Qwen3 Next Flash weights have been released on Hugging Face, with day-zero support provided by Unsloth.
In a browser OS visual recreation test, Bijan Bowen described Qwen3 Next Flash as the clearest image recreation he had seen to date.
In a C++ 3D racing game generation test, Qwen3 Next Flash generated a complete game featuring drifting, crashes, terrain elevation, spectators, a minimap, and a speedometer under the condition that the Raylib library was disabled.
Qwen3 Next Flash's performance in the 3D racing game test was comparable to GPT-5.2 Pro, with Qwen completing the task under the harder condition of having Raylib disabled.
In a 3D wrestling game task, Qwen3 Next Flash proactively chose to use Blender to create assets and Godot to build the game, rather than a Three.js approach.
Qwen models have a consistent weakness in spatial understanding, exhibiting scene flipping and incorrect object placement in 3D room rendering tests.
In spatial reconstruction accuracy for 3D room rendering, DeepSeek V4 Flash outperforms Qwen3 Next Flash.
1 more timeline events
All Facts (5)
The 4-bit quantized version (Q4K XL) of Qwen3 Next Flash is approximately 111 GB.
70%VerifiedThe 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%UnverifiedThe Ngram embedding table of Qwen3 Next Flash can be offloaded to host memory (CPU RAM) and executed with overlap with model computation via asynchronous prefetching.
50%UnverifiedQwen models have a consistent weakness in spatial understanding, exhibiting scene flipping and incorrect object placement in 3D room rendering tests.
50%UnverifiedQwen3 Next Flash weights have been released on Hugging Face, with day-zero support provided by Unsloth.
50%