Published: August 03, 2026
DeepSeek released DeepSeek-V4-Flash-0731 on July 31, 2026, publishing the weights on Hugging Face under an MIT licence and moving the official V4-Flash API into public beta. DeepSeek-V4-Flash-0731 is a 284 billion parameter Mixture-of-Experts model that activates 13 billion parameters per token and supports a 1 million token context window, priced at 0.14 dollars per million input tokens and 0.28 dollars per million output tokens. It is a re-post-trained version of the April preview rather than a new architecture, and DeepSeek reports that it beats the larger V4-Pro (Preview) on every agentic benchmark it published.
What did DeepSeek release with DeepSeek-V4-Flash-0731?
DeepSeek-V4-Flash-0731 is the official release of DeepSeek V4-Flash, superseding the preview version that has been available since April 2026. The model card states that the architecture and the parameter count are unchanged. The improvements come entirely from a new post-training pipeline focused on coding, agents, reasoning and tool use. The published checkpoint reports 304 billion parameters on Hugging Face, which includes the DSpark speculative decoding draft module sitting on top of the 284 billion parameter base model.
On the API side, the deepseek-v4-flash endpoint now natively supports the Responses API format and has been adapted for Codex. The V4-Pro API and the DeepSeek app and web models were not updated in this release. The weights are MIT licensed and ungated, so commercial on-premise deployment is permitted.
DeepSeek-V4-Flash-0731 benchmarks and technical specs
DeepSeek-V4-Flash-0731 is a decoder-only Mixture-of-Experts model with 284 billion total parameters and 13 billion activated per token. Each MoE layer holds 1 shared expert and 256 routed experts with an intermediate dimension of 2048, and 6 routed experts fire per token. The first three MoE layers use hash routing and multi-token prediction depth is 1. Attention combines Compressed Sparse Attention and Heavily Compressed Attention, and Manifold-Constrained Hyper-Connections replace conventional residual connections. Pre-training used more than 32 trillion tokens with the Muon optimizer.
On the benchmarks DeepSeek published with the model card, DeepSeek-V4-Flash-0731 scores 82.7 on Terminal Bench 2.1 against 61.8 for the V4-Flash preview and 72.1 for V4-Pro (Preview). It reaches 54.2 on NL2Repo, 76.7 on Cybergym, 54.4 on DeepSWE, 70.3 on Toolathlon-Verified, 25.2 on Agents’ Last Exam and 25.1 on AutomationBench Public. For comparison, DeepSeek lists GLM-5.2 at 81.0 on Terminal Bench 2.1 and Opus-4.8 at 85.0. All of these figures are vendor-reported and the code agent tasks were run with the minimal mode of the DeepSeek Harness, which has not been released, so independent runs may diverge.
How does DeepSeek-V4-Flash-0731 compare to DeepSeek V4-Pro?
The headline comparison in this release is internal. DeepSeek-V4-Flash-0731 scores higher than DeepSeek V4-Pro (Preview) on all nine agent and coding benchmarks the company published, while costing roughly a third as much on output tokens. V4-Flash output is priced at 0.28 dollars per million tokens against 0.87 dollars per million for deepseek-v4-pro. Input is 0.14 dollars per million tokens on a cache miss and 0.0028 dollars per million on a cache hit, with a concurrency limit of 2,500.
Against external models, DeepSeek positions V4-Flash-0731 close to GLM-5.2 on agentic coding and below Claude Opus 4.8, which leads the published table on Terminal Bench 2.1, NL2Repo, Cybergym, DeepSWE, Toolathlon-Verified, Agents’ Last Exam and AutomationBench Public. The practical argument is price-performance rather than a leaderboard win.
DeepSeek-V4-Flash-0731 availability, pricing and self-hosting
DeepSeek-V4-Flash-0731 is available through the DeepSeek API as deepseek-v4-flash, now in public beta, and as MIT-licensed weights on Hugging Face. API access requires no GPU capacity and is the practical route for most teams at 0.14 dollars per million input tokens and 0.28 dollars per million output tokens.
Self-hosting is a different matter. Every expert stays resident in memory even though only 13 billion parameters activate per token. DeepSeek’s own vLLM example serves the model on a single 4x GB300 node. Unsloth’s dynamic GGUF builds put the lossless 8-bit version at 162 GB and a 3-bit version at 103 GB, needing roughly 110 GB of combined RAM and VRAM. DSpark speculative decoding is enabled with a single vLLM flag and the DSpark paper reports 60 to 85 percent faster per-user generation on V4-Flash compared with the MTP-1 baseline at matched aggregate throughput. DeepSeek recommends a temperature of 1.0 and top_p of 0.95 for agentic use, with up to 384,000 output tokens at high and max reasoning effort.
Full details are in the DeepSeek-V4 model card on Hugging Face and in the DeepSeek API pricing documentation.
