DeepSeek released the official version of DeepSeek-V4-Flash, a 304-billion-parameter open-weight model, on August 2, putting the V4-Flash API into public beta with an upgrade focused on agent tasks1,2.
The model weights have been open-sourced and immediately reached second place on the HuggingFace trending chart, trailing only Kimi K3. DeepSeek says V4-Flash-0731 keeps the same structure and size as the preview version but has been retrained, with what the company describes as "substantially enhanced agentic capabilities"7.
Despite its 304 billion total parameters, V4-Flash activates only 13 billion parameters at inference time and occupies 167GB on Hugging Face5. Artificial Analysis ranks it ahead of MiniMax M3, a 428-billion-parameter model. DeepSeek says V4-Flash scored 82.7 on Terminal Bench 2.1 and 54.4 on DeepSWE. Independent testing on VulcanBench placed V4-Flash first, surpassing Claude Fable 5. The company says the model, after minimal fine-tuning, substantially outperforms the V4-Pro preview version and nearly matches Claude Opus-4.8 in capability.
Pricing is set at $0.14 per million input tokens and $0.27–$0.28 per million output tokens. Simon Willison noted the model "may currently be the best value-per-intelligence model out there".
The release adds support for the Responses API and is adapted for Codex. The update applies only to the V4-Flash API; the V4-Pro API and the models used on DeepSeek's app and website have not changed.
Alongside the model, DeepSeek introduced its Harness framework, a toolchain designed for structured evaluation of agentic software engineering tasks3. Harness provides standardized interfaces for tool-contract following, evidence-conclusion binding, and verification workflows. The V4-Flash official version was itself validated through Harness. DeepSeek is inviting open-source developers to join beta testing for Harness, according to a social media post by Cui Tianyi, who leads the effort.
ANALYSIS The Harness framework signals DeepSeek's expansion beyond pure model releases into AI coding infrastructure — building the evaluation and orchestration layer around its models, not just the models themselves.
The release follows a pattern of DeepSeek launches that have pressured pricing across the industry. The company's DeepSeek-R1, released earlier, topped OpenAI's o1 on AIME 2024, MATH-500, and SWE-Bench Verified benchmarks with open weights[1]. More broadly, Chinese open-weight models have been gaining traction globally; daily token volume for Zhipu's GLM-5.2 surged 50-fold on Vercel since mid-June as businesses shifted from premium US closed-source systems to cheaper alternatives[3].
V4-Flash output pricing is $0.27–$0.28 per million tokens. ANALYSIS That output price point, combined with open weights and benchmark performance competitive with Claude Opus-4.8, intensifies cost-performance pressure on closed-source providers. The 13-billion activated parameter count at inference, within a 304-billion total parameter model, underscores the efficiency gains DeepSeek is extracting from sparse architectures.
The release has been described across the industry as the "third DeepSeek moment," following DeepSeek-R1 in January 2025 and the Kimi K3 release cycle. DeepSeek has confirmed that a V4-Pro official version is in development.