DeepSeek V4 Flash 0731: The Small Model Goes Official First

DeepSeek V4 Flash 0731 is the official release of DeepSeek’s small agent model — announced July 31, 2026 in the company’s API changelog, which moves DeepSeek-V4-Flash out of Preview into public beta as a re-post-trained checkpoint with the exact same 284B-total / 13B-active architecture it launched with in April.

The release matters for a reason most same-day coverage skipped: the sequencing. DeepSeek shipped the official version of its smaller model before the 1.6T-parameter V4-Pro flagship, and its changelog markets the small model’s agent-benchmark scores as far exceeding the bigger one’s — on DeepSeek’s own harness. Meanwhile the announced peak/off-peak pricing isn’t live, and the -0731 weights haven’t appeared on Hugging Face.

This guide separates the confirmed facts from the vendor claims: what actually changed between Preview and 0731, how the vendor-stated benchmark numbers should be read, what the blended cache-hit price math looks like for agent workloads, and why the 2,500-request concurrency ceiling may be the most operationally significant line in the announcement.

The announcement itself is narrow and precise. DeepSeek’s changelog states that V4-Flash exits Preview into official release — public beta — as a checkpoint named DeepSeek-V4-Flash-0731. The company is explicit that this is not a new model: same architecture, same size, same 284B total / 13B activated Mixture-of-Experts configuration as the April 24 Preview, with only the post-training stage redone.

The scope is equally narrow: the update applies to the V4-Flash API only. DeepSeek’s own app, website, and the V4-Pro API were not touched by this release. Independent corroboration arrived the same day from at least two outlets — TechNode (“DeepSeek puts V4-Flash API into public beta”) and an Odaily newsflash carrying the translated framing that only retraining was conducted on the post-training stage. DeepSeek’s own X account announced the release the same day, headlining “massively upgraded Agent capabilities.”

Specs carried over unchanged from the Preview: a 1M-token context window and a 384K-token maximum output, with both thinking (default) and non-thinking modes supported. If the architecture is new to you, our guide to the V4 architecture covers the design this release deliberately leaves alone.

Most coverage reported “V4-Flash went to public beta” and moved on. The more interesting reading is what DeepSeek chose to productionize first. The 284B / 13B model — the cheap one, the one with 5× the concurrency ceiling — reached official status while the 1.6T flagship stayed in Preview. And DeepSeek’s changelog frames the 0731 checkpoint’s agent-suite scores as far exceeding V4-Pro-Preview’s. That is the vendor’s own comparative claim about its own two models on its own benchmark suite, not an independent finding — but as a statement of intent it’s unambiguous: DeepSeek is positioning the small model as its agent workhorse.

Our interpretation: this tracks a broader 2026 pattern where labs increasingly treat agentic workloads — high-volume, tool-calling, parallel — as the product, and route them to small models where post-training quality matters more than parameter count. A re-post-train-only release that claims large agent gains is effectively an argument that the remaining frontier in agent performance lives in post-training, not scale. If the claimed numbers hold up under third-party evaluation, that argument gets a lot stronger; if they don’t, the release is still a meaningful price-performance event because the economics below don’t depend on the benchmarks at all.

The headline numbers below come from DeepSeek’s own changelog, produced on DeepSeek’s own evaluation harness, reported as run in minimal mode at maximum reasoning effort. No independent lab had reproduced any of these figures as of July 31, 2026. Read every bar in this chart as a vendor claim awaiting verification, not a measured fact — most secondary coverage repeated the numbers without that caveat, which is exactly why we’re leading with it.

There is one semi-independent data point available. Artificial Analysis lists “DeepSeek V4 Flash 0731” with an Intelligence Index of 50 — which it describes as well above average among comparable models, against a stated median of 17. That is a genuinely non-vendor measurement, but note what it isn’t: AA’s index measures general intelligence, not the agent-suite tasks DeepSeek’s own numbers describe. Treat it as directional corroboration that the model is strong, not as validation of any specific score above.

No outlet we reviewed laid out a single side-by-side view of what actually changed between the April Preview, the 0731 official release, and the still-Preview V4-Pro. Here’s that ledger, compiled July 31, 2026 from DeepSeek’s changelog, its pricing page, and Hugging Face.

A dash means the attribute isn’t documented for that checkpoint in DeepSeek’s current public docs (the pricing page describes the live models, not the retired Preview configuration). The single most consequential row for most teams is the last one: the official release is, as of today, an API-only artifact.

Official 0731 pricing per 1M tokens: $0.0028 cache-hit input, $0.14 cache-miss input, $0.28 output. V4-Pro’s unchanged Preview rates are $0.003625 / $0.435 / $0.87. On cache-miss input and output, Flash costs roughly one-third of Pro — about 3.1× cheaper on both (our calculation from DeepSeek’s listed prices: $0.435 ÷ $0.14 and $0.87 ÷ $0.28 both equal ~3.1).

Sticker prices don’t tell an agent operator much, though, because agent loops replay long shared prefixes and cache-hit rates vary wildly by workload shape. So we modeled the blended input cost at three cache-hit rates — the percentage of your input tokens billed at the cache-hit price. Formula per cell: blended input cost = (cache-hit share × cache-hit price) + (cache-miss share × cache-miss price). This is our own arithmetic from DeepSeek’s listed prices, not a DeepSeek-published table.

Two readings. First, output tokens are unaffected by caching, so the output-side ratio is constant at 3.11× ($0.87 ÷ $0.28) — and Artificial Analysis’s verbosity note above (210M output tokens in its suite vs a 62M median) means the output line may dominate real agent bills. Second, the Flash advantage on input narrows slightly as cache-hit rates rise — at a 90% hit rate the input gap is 2.83× rather than 3.1×, because Pro’s cache-hit price is proportionally closer to Flash’s ($0.003625 vs $0.0028, about 1.3×). Cache-heavy architectures shrink the sticker-price story; they don’t erase it. For why a $0.28/1M output price is even viable, see DeepSeek’s open-weight economics.

Most coverage buried the line that matters most to anyone running agents at scale: deepseek-v4-flash supports 2,500 concurrent requests versus deepseek-v4-pro’s 500 — a 5× headroom gap, per DeepSeek’s own model documentation. Benchmark scores are claims; concurrency limits are hard operational constraints you hit on day one. A team running a fleet of parallel agent sessions — bulk code migration, wide research fans, high-volume document processing — is throttled by that ceiling long before model quality differences show up.

This is the release’s most defensible value proposition precisely because it doesn’t depend on contested numbers. Even a skeptic who discounts every benchmark in Section 03 still faces the same routing question: if your workload is wide and parallel, the official Flash endpoint offers materially more capacity per account at materially lower cost than the flagship.

The 0731 release also carries a compatibility story. Per the changelog, the official V4-Flash natively supports OpenAI’s Responses API format, specifically adapted for Codex — meaning Codex-style agent scaffolds can point at the Flash endpoint without rewriting their request layer. Pro users wait: DeepSeek’s docs state “We will add support for the deepseek-v4-pro model in early August 2026” — a DeepSeek-stated target, not a guarantee.

Alongside that, both V4-Flash and V4-Pro support the Anthropic API message format, reachable at https://api.deepseek.com/anthropic with the same model IDs — confirmed live alongside this release rather than new today. We unpack what this two-format convergence means for multi-vendor agent stacks in today’s companion piece on DeepSeek’s Responses API and Anthropic-format support.

Here is the gap most surface coverage didn’t raise. As of July 31, 2026, no -0731-suffixed repository exists on Hugging Face. The DeepSeek-V4-Flash repos that do exist — deepseek-ai/DeepSeek-V4-Flash, DeepSeek-V4-Flash-Base, and DeepSeek-V4-Flash-DSpark — are the April Preview generation; none references “0731” anywhere on its model card. The existing Preview repo is MIT-licensed and lists 284B parameters, so the open-weight lineage is real. But the official release, today, is an API-only artifact.

Whether that changes is a genuinely open question, and we’d frame it exactly that way rather than predict timing. The pattern-consistent case: DeepSeek open-sourced the original V4 Preview on Hugging Face on day one (April 24, 2026), so a follow-up weight drop for 0731 would fit its established behavior. The skeptical case: an official, agent-marketed, revenue-carrying endpoint is precisely the kind of asset a lab might keep API-only. Teams running the open-weight stack — the audience of our V3.2-to-V4 migration playbook — should watch the Hugging Face org rather than assume parity with the API checkpoint.

The decision tree splits cleanly by workload shape and by how much you need the claims to be true before acting.

Looking forward, the projection we’re comfortable making: expect a near-term wave of third-party agent-benchmark runs against the 0731 endpoint, and expect the interesting divergence to be between DeepSeek’s own harness numbers and independent harnesses with different tool stacks — agent scores are notoriously harness-sensitive, which cuts both ways. If independent runs land anywhere near the vendor’s figures, the price-per-capability pressure on mid-tier closed models intensifies sharply going into Q4. If your team wants help designing that evaluation — routing real workloads across candidate endpoints and measuring cost per completed task rather than leaderboard deltas — that’s the first step of our AI transformation engagements.