DeepSeek Open-Sources V4-Flash Model with Million-Token Context

On July 31, DeepSeek unexpectedly released the weights and technical report for its latest model, DeepSeek-V4-Flash-0731, on Hugging Face. Featuring million-level token context intelligence and impressive benchmark performance, the model has quickly captured the community's attention.

Confirmed

• The model has a total parameter count of 304B, natively supports million-level token context, and includes a speculative decoding module.
• It supports 8-bit and fp8 precision, is built on the Transformers architecture and Safetensors format, and is released under the MIT license.
• The official statement notes that this version replaces the previous preview. Although it has significantly fewer activated parameters than the V4-Pro preview, it substantially enhances agentic capabilities.

Why it matters

• According to tester @tomaarsen, this 304B parameter model outperforms the 750B parameter GLM5.2, demonstrating exceptional parameter efficiency.
• This open-source release equips local deployment and community developers with cutting-edge tools. Hugging Face Director Victor Mustar expressed amazement at the rapid release, and the community expects local quantized formats like GGUF to follow shortly.

2026-07-31 ~ 2026-07-31 · 8 related posts

• [source] DeepSeek Releases New DeepSeek-V4-Flash-0731 Model — deepseek-ai · 2026-07-31
• DeepSeek-V4-Flash-0731 Model Weights Officially Released — shing3232 · 2026-07-31
• DeepSeek-V4-Flash Surfaces on Hugging Face with Million-Token Context — victormustar · 2026-07-31
• [source] DeepSeek-V4-Flash Released: 304B Model Beats 750B GLM5.2 — tomaarsen · 2026-07-31
• [source] DeepSeek Open-Sources V4 Flash 0731 Weights with Tech Report — kimmonismus · 2026-07-31
• DeepSeek V4 Flash Weights Released on Hugging Face — gaganghotra_ · 2026-07-31
• DeepSeek-V4-Flash Open Source Sparks Debate: Open Weights Getting Highly Competitive — eyishazyer · 2026-07-31

1 near-duplicate retellings: _akhaliq