DeepSeek Releases V4-Flash-0731 Model with Agentic and Coding Enhancements

DeepSeek published DeepSeek-V4-Flash-0731 on Hugging Face and simultaneously moved the official V4-Flash API into public beta on July 31, 2026. This release represents the official version, superseding a previous preview, with no changes to the underlying architecture or model size. The performance improvements stem from re-post-training rather than a new design. The model checkpoint includes the DSpark speculative decoding module, aligning with the structure of DeepSeek-V4-Flash-DSpark. Hugging Face lists the repository's parameter count at 304 billion, which accounts for the DSpark module layered on top of the 284 billion base parameters. On the API front, the deepseek-v4-flash model now natively supports the Responses API format and has been adapted for coding tasks, specifically referencing Codex. Notably, the V4-Pro API, along with the associated app and web models, did not receive updates in this release. Deployment of DeepSeek-V4-Flash-0731 is feasible through two distinct methods. Via API, DeepSeek offers pricing at $0.14 per 1 million input tokens during a cache miss and $0.0028 per 1 million input tokens on a cache hit. Output tokens are priced at $0.28 per 1 million, with a concurrency limit of 2,500. This API pricing is approximately one-third of the output cost for deepseek-v4-pro, which is listed at $0.87 per 1 million tokens. This pricing structure enables seed-stage startups, independent developers, and internal platform teams to run agent loops without requiring a dedicated GPU budget. Self-hosting presents a higher barrier to entry. While the model weights are available under an MIT license and are ungated, all experts within the model remain resident in memory, even though only 13 billion parameters are activated per token. DeepSeek's provided vLLM example demonstrates serving the model on a single 4xGB300 node. Using Unsloth's dynamic GGUFs, a lossless 8-bit build requires 162 GB of storage, and a 3-bit build requires 103 GB, necessitating approximately 110 GB of combined RAM and VRAM. Self-hosting is therefore best suited for mid-size to large enterprises equipped with a serving cluster or a single, high-performance workstation capable of aggressive quantization. According to the DeepSeek-V4 technical report, V4-Flash is a Mixture-of-Experts (MoE) model with 284 billion parameters. It activates 13 billion parameters per token and features a context window of 1 million tokens. Each MoE layer incorporates one shared expert and 256 routed experts, with an intermediate dimension of 2048. Six routed experts are activated for each token. The initial three MoE layers are detailed in the technical report.