Official API Beta Launch
DeepSeek officially transitioned its V4-Flash model to public beta on July 31, 2026, releasing the updated DeepSeek-V4-Flash-0731 build. This release marks the transition from preview status to an official product, bringing significant performance improvements in coding and autonomous agent tasks. According to the official model card, the underlying architecture remains identical to the April preview, featuring a 284B-parameter Mixture-of-Experts (MoE) design with 13B parameters activated per token and a 1-million-token context window.
Performance Gains Through Post-Training
The performance jump observed in the 0731 release is attributed to intensive re-post-training rather than architectural changes. Benchmark results reported by DeepSeek show substantial improvements across key categories. For instance, the model achieved a score of 76.7 on the Cybergym benchmark, compared to 38.7 in the preview version. Similarly, in the DeepSWE coding benchmark, the model climbed from 7.3 to 54.4. These gains suggest that the model is now significantly more capable of handling complex software engineering workflows and multi-step agentic reasoning.
Deployment and Accessibility
DeepSeek-V4-Flash-0731 is now available via API at a competitive price point of $0.14 per 1M input tokens on a cache miss, making it roughly one-third the cost of the V4-Pro output pricing. This accessibility is designed to support startups and developers in running agentic loops without high infrastructure overhead. For enterprises requiring on-premise solutions, the model is MIT-licensed and ungated. However, self-hosting remains resource-intensive; full-precision serving requires a 4×GB300 node, while aggressive 3-bit quantization can operate on approximately 110 GB of combined memory.
Technical Integration
The model includes the DSpark speculative decoding module, which can be enabled in vLLM environments to achieve 60–85% faster generation speeds. Unlike standard models that utilize Jinja chat templates, DeepSeek has implemented custom encode_messages and parse_message_from_completion_text methods. The company recommends a temperature of 1.0 for agentic use, with support for up to 384K output tokens when using the ‘high’ or ‘max’ reasoning effort settings.