DeepSeek-V4-Flash Released: 304B Model Beats 750B GLM5.2

tomaarsen · x · 2026-07-31

DeepSeek has released DeepSeek-V4-Flash-0731 on Hugging Face. The 304B parameter model supports million-token context and utilizes 8-bit and fp8 precision.

According to tester @tomaarsen, the final checkpoint shows a massive performance gap, surprisingly outperforming the previous V4-Pro preview. Furthermore, it manages to beat the 750B parameter GLM5.2 model, setting high expectations for the eventual release of DeepSeek-V4-Pro.

Related event: DeepSeek Open-Sources V4-Flash Model with Million-Token Context(8 posts)→

Original post →

More from Models

Models channel →