DeepSeek quietly turned its budget model into an agentic benchmark leader on July 31, posting a new V4-Flash-0731 checkpoint to Hugging Face and moving the V4-Flash API into public beta the same day. The architecture was left untouched; all the gains come from a post-training cycle, yet the small model now beats the company’s own Pro flagship on every agentic benchmark DeepSeek published.
The headline results show a striking jump from the preview build. On Terminal Bench 2.1, the measure developers use for agentic coding tasks, V4-Flash-0731 scores 82.7, up from 61.8 for the preview and ahead of V4-Pro’s 72.1. On DeepSWE, a software engineering benchmark, the model improved from 7.3 to 54.4 in a single refresh. Cybergym, which tests security-agent performance, climbed from 38.7 to 76.7. The improvement appears targeted: the same post-training cycle left the model broadly competitive with GLM-5.2 and Anthropic’s Opus-4.8 on several agentic tests while remaining far smaller at inference time.
Those gains come with caveats. The benchmark figures are vendor-reported and were produced using the minimal mode of DeepSeek’s internal Harness framework, which has not been released, meaning independent evaluation could diverge. Two additional test sets, DSBench-FullStack and DSBench-Hard, are internal to DeepSeek. The company’s own documentation flags that agent scores are sensitive to the harness used.
The pricing is where the release gets interesting for developers. The API charges $0.14 per million input tokens on a cache miss, $0.0028 per million on a cache hit, and $0.28 per million output tokens, roughly a third of V4-Pro’s output rate. The model keeps the 284-billion-parameter mixture-of-experts design with 13 billion parameters active per token, a 1-million-token context window, and up to 384,000 output tokens at high reasoning effort. The checkpoint ships with DeepSeek’s DSpark speculative decoding module attached, which the company says delivers 60-85 percent faster per-user generation in matched-throughput serving.
The API upgrade path is deliberately frictionless: developers switching from the preview need only change the model identifier in their client code. The new build natively supports the Responses API format and is adapted for Codex, positioning it for the agentic and coding workloads DeepSeek appears to expect will dominate traffic. The V4-Pro API and app models were not updated in this release.
For teams that want to avoid the hosted API entirely, the weights are MIT-licensed and ungated, enabling on-premise deployment. That path is not trivial: serving at full precision calls for a four-node GB300 cluster in DeepSeek’s vLLM reference configuration, while aggressive quantization brings requirements down to roughly 110 GB of combined RAM and VRAM, with Unsloth’s lossless 8-bit build at about 162 GB and a 3-bit build at about 103 GB.
The strategic signal is the model ladder: DeepSeek is steering everyday agentic and coding traffic toward Flash at a third of the price while reserving its hardest reasoning workloads for Pro. The refresh hands startups and independent developers near-flagship agent capability at a fraction of the cost of equivalent proprietary models.
Sources: DeepSeek Upgrades DeepSeek-V4-Flash-0731 with Major Agentic and Coding Gains (MarkTechPost, July 31, 2026); DeepSeek-V4-Flash-0731 model card (Hugging Face); DeepSeek puts V4-Flash API into public beta (TechNode, July 31, 2026); DeepSeek API pricing (DeepSeek)