डीपसीक V4 फ्लॅश

The earlier V4 Flash release: the same 284B-total, 13B-active Mixture-of-Experts design and million-token context, designed for fast inference at scale. It remains one of the cheapest genuinely capable models available and supports non-think, high and max reasoning settings. The 0731 revision supersedes it with better agent tuning.

This is the checkpoint that opened the V4 generation. DeepSeek launched V4-Pro and V4-Flash together as a preview, open-sourced both, published a technical report, and framed the whole release around one idea: making a million tokens of context cheap enough to use by default. It went further than a claim - a million tokens became the standard window across every official DeepSeek service on the day, and Flash was the tier built to serve it at volume, appearing in the chat product as Instant Mode.

DeepSeek positions Flash against its own flagship rather than against the field. It says the smaller model's reasoning comes close to V4-Pro, that the two perform on a par on simple agent tasks, and that Flash trades the remaining gap for smaller size, faster responses and a much lower price. Its published table for the preview weights at max effort puts Flash at 86.2 on MMLU-Pro, 88.1 on GPQA Diamond, 34.8 on Humanity's Last Exam and 91.6 on LiveCodeBench, with 78.7 on the million-token MRCR retrieval test and 79.0 on SWE-bench Verified - close to Pro on knowledge and reasoning, further behind on the agentic column, where Pro reached 67.9 on Terminal Bench 2.0 against Flash's 56.9.

The efficiency story is architectural. DeepSeek describes a new attention design that pairs token-wise compression with its own sparse attention, and credits it for long-context handling at sharply reduced compute and memory cost - the reason a million-token window could be made the default rather than a premium tier. The lab also says V4 was tuned deliberately for agent tools, calls out integration with the popular coding agents of the day, and notes it was already driving DeepSeek's own in-house agentic coding.

Two things date this checkpoint. The release set a retirement clock on the legacy `deepseek-chat` and `deepseek-reasoner` names, which pointed at Flash's non-thinking and thinking modes until they were switched off in July 2026. And Flash itself was superseded three months later by a re-post-trained revision that kept the same architecture and weights count but scored far higher on agent work - so this build is best understood as the cheap, capable baseline of the V4 line rather than its current form.