Reviews · JULY 31, 2026
DeepSeek-V4-Flash-0731 exits preview: re-post-trained 284B MoE beats its own 1.6T flagship on agent benchmarks
DeepSeek's July 31 official release keeps the preview architecture unchanged but a full re-post-training pass lifts vendor-stated Terminal-Bench 2.1 from 61.8 to 82.7 — while Artificial Analysis flags 3.4× median verbosity behind the numbers.
DeepSeek pushed its V4-Flash API from preview into public beta on July 31, 2026, shipping the build as DeepSeek-V4-Flash-0731 with the same architecture as the April preview and, per the change log, "only re-post-trained." The company's headline claim is that this 284-billion-parameter MoE now beats its own still-preview 1.6T V4-Pro flagship on agent evaluations, showing "significantly enhanced agent capabilities, with benchmark results far exceeding V4-Pro-Preview."
The numbers make the case aggressively. Terminal-Bench 2.1 climbs from 61.8 to 82.7 between the preview and the 0731 build, and DeepSeek's Hugging Face card asserts the model "outperforms DeepSeek-V4-Pro (Preview) on benchmarks listed below despite its far smaller activated parameter count," pointing at DSWE-Bench and DSBench-Hard alongside Terminal-Bench. Runs used DeepSeek Harness in "minimal mode" at max reasoning effort, with temperature 1.0 and top_p 0.95.
Independent scoring largely corroborates the vibe shift, with an asterisk. Artificial Analysis puts the max-effort build at 50 on its Intelligence Index versus a median of 17 in the same price tier, over a 1M-token context. It also logs the token bill. Evaluating the index consumed 210M output tokens against a 62M median for reasoning models, a 3.4× verbosity ratio, for a total eval cost of $72.02. Blended pricing on Artificial Analysis's 7:2:1 cache-hit/input/output mix works out to $0.06 per 1M tokens; output-heavy reasoning pays the headline $0.28 per 1M.
That gap between blended and output-heavy pricing is where the review lives. Flash is genuinely cheap on paper. Turn max reasoning on for agentic workloads, where the 384K-token output cap and the 7-token DSpark speculative-decoding setting are doing real work, and the effective bill drifts toward the top of the tier. The intelligence is real; so is the verbosity tax.
The deployment surface got the more interesting update. The 0731 build adds native Responses API support and is, per DeepSeek, "specifically adapted for Codex," which reads as a direct pitch at the agent-tooling ecosystem OpenAI has cultivated. Local operators get vllm serve, an FP8 KV cache, and the deep_gemm_mega_moe backend for expert parallelism, all under an MIT license on the weights. Recommended settings split by use case: top_p 0.95 for agentic runs, 1.0 otherwise.
V4-Pro's API remains unchanged, with an official release that "will follow soon." Bloomberg framed the July 31 move as the public beta of DeepSeek's flagship API, which is technically correct and structurally misleading. The flagship, by parameter count, is still Pro. The flagship, by benchmark, is now the smaller model DeepSeek chose to ship first.