DeepSeek V4-Flash Official Release: Open‑Source Model Outperforms V4‑Pro Preview

DeepSeek V4‑Flash release

DeepSeek released the 304‑billion‑parameter V4‑Flash model. The model was only fine‑tuned after pre‑training, yet its inference performance exceeds the V4‑Pro preview and is comparable to Claude Opus‑4.8.

Open‑source weights and community ranking

Model weights are publicly available on HuggingFace (https://huggingface.co/deepseek-ai/DeepSeek-V4-Flash-0731). The model currently ranks second on HuggingFace’s trend list, behind Kimi K3.

Cost‑performance metrics

Community testing reports a performance score of 82.7 with a token price of $0.14 / $0.28, giving a cost‑performance advantage over leading closed‑source models.

Benchmark results

Independent evaluation on VulcanBench shows DeepSeek V4‑Flash outperforming Claude Fable5 and achieving first place on the benchmark.

Community reaction

Users refer to this as the third “DeepSeek moment,” following the earlier R1 and K3 releases.

Future releases

The V4‑Pro official version is announced as forthcoming.

Harness framework and AI coding outlook

DeepSeek is developing a Harness framework for an AI coding assistant; the V4‑Flash analysis was performed with this framework. The community anticipates a forthcoming DeepSeek Code product, which could impact the AI coding market.

Reference: https://x.com/deepseek_ai/status/2083084415157022911