The same model, retrained
DeepSeek has moved V4-Flash out of preview and into general release under the build name DeepSeek-V4-Flash-0731, dated to its July 31, 2026 release. The notable part isn't new hardware or a bigger network: it is the same 284-billion-parameter mixture-of-experts (MoE) design as the April preview. A mixture-of-experts model splits its parameters into many specialized sub-networks, or experts, and routes each token to only a handful of them instead of running every parameter on every token; here, that means 13 billion of the model's 284 billion total parameters are active on any given token, split across one shared expert plus 256 routed experts per layer. The model also keeps the same hybrid attention setup from the preview, pairing Compressed Sparse Attention with Heavily Compressed Attention, two techniques for cutting the computational cost of tracking relationships across a long input. Nothing about that architecture changed between preview and release. What changed is the post-training pipeline that sits on top of it, the later stage of training where a base model is refined on curated tasks and feedback rather than raw text.
That retraining alone was enough to push V4-Flash-0731 past DeepSeek's own larger and still-in-preview V4-Pro model, which runs 1.6 trillion total parameters with 49 billion active, on every one of the nine agentic benchmarks the company publishes. A model activating roughly a quarter of the parameters per token is now the better agent, at least by DeepSeek's own numbers.
What the numbers say
DeepSeek's published scores for V4-Flash-0731 include 82.7 on Terminal-Bench 2.1 (a benchmark of real command-line task completion), 54.4 on DeepSWE, and 70.3 on Toolathlon-Verified, alongside gains on NL2Repo and Cybergym. All of these are vendor-reported on DeepSeek's own evaluation harness, which has not yet been independently reproduced, so they should be read as a company's claim about its own model rather than a settled result. One outside data point exists: independent evaluator Artificial Analysis put the model's general Intelligence Index at 50, a measure of broad reasoning rather than agentic task completion specifically, so it isn't a direct check on DeepSeek's agent numbers.
The model also ships with an optional add-on called DSpark, a speculative-decoding draft module that pairs a smaller predictor model with the 284B base (a combined 304B when both are loaded) to speed up generation without changing the base model's outputs. Speculative decoding works by having the small draft module guess several tokens ahead, which the full model then checks in one pass instead of generating each token from scratch, and DSpark is offered as an opt-in rather than something bundled into every request.
Context window and pricing
V4-Flash-0731 takes up to 1 million tokens of input and can generate up to 384,000 tokens of output in a single response, unchanged from the preview. Pricing did move: DeepSeek is charging $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 through its API. That works out to roughly a third of what V4-Pro costs to run. The API also now accepts requests in the same Responses API format OpenAI's Codex tooling uses, and DeepSeek raised the concurrent-request ceiling to 2,500, five times V4-Pro's limit of 500. Separately, DeepSeek has announced plans to introduce peak-hour pricing that would double all billable rates for seven hours a day, though that surcharge was not active at release.
What it means in practice
For teams building on DeepSeek's API, the practical upshot is that V4-Flash-0731 is now the model to default to for agentic and coding workloads rather than a cheaper fallback: it costs less to run than V4-Pro and DeepSeek's own numbers say it performs better on the tasks that matter for tool-using agents. Anyone relying on those benchmark claims for a purchasing or architecture decision should still run their own evaluation, since the published figures come from DeepSeek's harness and have not been checked by an outside party. The bigger signal for the field is that DeepSeek got most of this jump from re-training rather than re-architecting, which is a cheaper lever than it looks and one competitors can pull too.
