DeepSeek V4 Flash
DeepSeek upgraded the deepseek-v4-flash API to the DeepSeek-V4-Flash-0731 public beta on July 31, 2026. The API name is unchanged and only the API post-training was updated; architecture remains a 284B-total/13B-active MoE with a 1M context window and 384K maximum output. The release improves coding and agent performance, supports the Responses API, tool calls, JSON output, and OpenAI/Anthropic-compatible interfaces. Current prices per 1M tokens are $0.0028 cache-hit input, $0.14 cache-miss input, and $0.28 output. The app/web models and open weights were not updated to 0731.
Data sourced primarily from official releases (GitHub, Hugging Face, papers), then benchmark leaderboards, then third-party evaluators. Learn about our data methodology
| Type | Condition | Input | Output |
|---|---|---|---|
| Text | - | $0.140/ 1M | $0.280/ 1M |
| Type | TTL | Write | Read |
|---|---|---|---|
| Text | - | - | $0.0028/ 1M |
DeepSeek-V4-Flash currently shows benchmark results led by LiveCodeBench (4 / 123, score 91.60), MMLU Pro (16 / 132, score 86.40), GPQA Diamond (31 / 187, score 88.10). This page also consolidates core specs, context limits, and API pricing so you can evaluate the model from benchmark results and deployment constraints together.
No curated comparisons for this model yet.
Want a custom combination? Open the compare tool
According to the official DeepSeek change log, DeepSeek upgraded the deepseek-v4-flash API to DeepSeek-V4-Flash-0731 on July 31, 2026 and opened it as a public beta. The model name and base URL are unchanged. This is an API-only post-training update focused on coding and agent capabilities; the architecture and parameter counts are unchanged, and the DeepSeek app/web models and open weights were not replaced by this update.
DeepSeek reports the following 0731 agent scores: Terminal Bench 2.1 82.7, NL2Repo 54.2, CyberGym 76.7, DeepSWE 54.4, Toolathlon Verified 70.3, Agents' Last Exam 25.2, and Automation Bench (Public) 25.1. The internal DSBench-FullStack and DSBench-Hard results are 68.7 and 59.6. DeepSeek states that the public Code Agent evaluations used its forthcoming minimal harness, max effort, temperature 1.0, and top_p 0.95.
DeepSeek-V4-Flash is a 284B-total, 13B-active MoE model with a 1M-token context window and up to 384K output tokens. The V4 family uses a hybrid attention architecture combining Compressed Sparse Attention and Heavily Compressed Attention, together with mHC connections and the Muon optimizer. The official open-weight checkpoint uses mixed FP4 expert and mostly FP8 non-expert parameters.
The model supports Non-think, Think High, and Think Max modes, with thinking enabled by default in the API. Official interfaces support OpenAI Chat Completions, the Anthropic API, the Responses API, JSON output, tool calls, prefix completion, and FIM completion in non-thinking mode. It accepts text and returns text; image input is not supported.
The official DeepSeek pricing page currently lists $0.0028 per one million cache-hit input tokens, $0.14 per one million cache-miss input tokens, and $0.28 per one million output tokens. DeepSeek has announced future peak pricing at twice the regular rate, but no effective date has been published, so that pending rule is not included in the active price table.
The April 24, 2026 DeepSeek-V4-Flash weights remain available from the official DeepSeek Hugging Face page under the MIT License. DeepSeek explicitly states that the 0731 update applies only to the API, so the existing open weights should not be represented as the 0731 checkpoint.
The current DeepSeek API version is DeepSeek-V4-Flash-0731 public beta, still called with deepseek-v4-flash. The 0731 update changes only the API post-training; the app/web models and April open weights were not replaced.
The official model is text input and text output only. It does not support image input.
DeepSeek-V4-Flash is a 284B-total, 13B-active MoE model with a 1M-token context window and up to 384K output tokens.
The architecture and size are unchanged. DeepSeek re-post-trained the API model to improve coding and agent capabilities, reporting scores such as 82.7 on Terminal Bench 2.1 and 54.2 on NL2Repo.
Current prices per 1M tokens are $0.0028 for cache-hit input, $0.14 for cache-miss input, and $0.28 for output. Announced 2x peak pricing is not yet active because no effective date has been published.
No. The April DeepSeek-V4-Flash weights are available under the MIT License, but DeepSeek says the 0731 update applies only to the API.
Follow DataLearner on WeChat for AI model updates and research notes.
