DeepSeek Upgrades V4-Flash-0731: Major Gains in Agentic Coding and API Beta

DeepSeek published DeepSeek-V4-Flash-0731 on Hugging Face and moved the official V4-Flash API to public beta on July 31, 2026. The model card states this is the official release superseding the preview, with unchanged architecture and size; gains come from re-post-training. The checkpoint includes the DSpark speculative decoding module, and Hugging Face reports 304B parameters including the draft module on top of the 284B base. The API now natively supports the Responses API format and is adapted for Codex, but V4-Pro API and app/web models were not updated. Deployment is possible via API at a low cost (about a third of the pro output price) or via self-hosting, which requires significant resources: MIT-licensed weights, but every expert stays resident in memory. A vLLM example serves it on a 4×GB300 node; Unsloth dynamic GGUFs fit in about 110GB memory at aggressive quantization. Architecture includes 284B parameters with 13B active per token, 1M-token context, hybrid attention with CSA and HCA, and manifold-constrained hyper-connections. Benchmarks show V4-Flash-0731 outperforming V4-Pro (Preview) on all agentic benchmarks, but all figures are DeepSeek-reported on an unreleased harness. DSpark is enabled with a vLLM flag, and the model uses an encoding folder instead of a Jinja chat template.