डीपसीक V4 फ्लैश 0731

The lighter, high-throughput tier of the DeepSeek V4 pair: a Mixture-of-Experts model with 284B total and 13B activated parameters and a one-million-token context. It uses a hybrid attention architecture combining Compressed Sparse Attention and Heavily Compressed Attention to make long context efficient. This re-post-trained revision is tuned for coding, reasoning and agent workflows, and offers non-think plus high and max thinking settings.

This is the checkpoint that took DeepSeek-V4-Flash out of preview. DeepSeek is unusually precise about what changed: the architecture and the parameter count are identical to the preview weights, and the entire difference is a fresh round of post-training aimed at agent work. Nothing about the model got bigger, and the price did not move; what moved is how reliably it finishes a task once a harness hands it tools.

The gain the lab reports is large enough that it reframes the tier. DeepSeek measures the 0731 revision at 82.7 on Terminal Bench 2.1, 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 the public AutomationBench split, plus 68.7 and 59.6 on DSBench-FullStack and DSBench-Hard, its own internal full-stack and hard-coding-agent sets. Its point is the comparison against its own previous flagship rather than against anyone else: on every one of those the cheap Flash checkpoint clears what V4-Pro-Preview scored. The lab is explicit that these were run through DeepSeek Harness in minimal mode at the max effort level, with top-p 0.95 and temperature 1.0, so they are peak-effort numbers, not defaults.

The release also made the model easier to drop into an existing agent. DeepSeek added native support for the OpenAI Responses API format alongside its OpenAI- and Anthropic-compatible chat endpoints, with a configuration path aimed specifically at Codex, and the calling name stayed `deepseek-v4-flash` so nothing downstream had to be rewritten. A fortnight later the lab exposed a low / high / max thinking-effort control across both V4 models, and recommends high for everyday agent loops and max only for genuinely hard scenarios.

Where it stops is a matter of tier rather than defect. DeepSeek shipped this update to the Flash API alone - the Pro API and the app and web experiences were untouched that day - and it published the Pro GA two weeks later with materially higher scores on the same evaluations. Flash is the model to reach for when throughput and cost dominate; the lab points at Pro for the hardest reasoning and the longest agent runs.