DeepSeek Releases V4 Flash Model for Highly Efficient Million-Token Context

MaziyarPanahi · x · 2026-07-31

DeepSeek has released the new DeepSeek-V4-Flash-0731 model on Hugging Face. According to the model card, it focuses on highly efficient million-token context intelligence, supports 8-bit and fp8 precision, and is open-sourced under the MIT license. The tweet mentions it achieved amazing scores on the Terminal Bench 2.1 evaluation.

Related event: DeepSeek Surprise-Releases V4-Flash Model with 1M Token Context(13 posts)→

Original post →

More from Models

Models channel →