DeepSeek V4-Pro GA and V4-Flash Explained

DeepSeek finished the V4 rollout in two moves, two weeks apart. On July 31, 2026 it published DeepSeek-V4-Flash-0731 as MIT-licensed open weights and put the matching API into public beta. On August 13, 2026 it took V4-Pro out of preview with a post titled "DeepSeek-V4-Pro GA Release", and changed how the whole family is priced three days later.

If you have seen both dates collapsed into a single "V4 GA on July 31" headline, that is wrong in a way that matters for anyone planning a migration. Flash and Pro are on different release tracks, different licences, and different access surfaces.

Key Dates

What Is DeepSeek V4?

DeepSeek V4 is a family of Mixture-of-Experts language models with a 1M-token context window, shipped in two sizes: V4-Pro, the 1.6T-parameter flagship, and V4-Flash, the 284B-parameter fast tier aimed at coding and agent work.

The efficiency claim is the reason the family got attention. The DeepSeek-V4 paper on arXiv reports that at a one-million-token context, V4-Pro needs 27% of the single-token inference FLOPs and 10% of the KV cache of the comparison baseline. That is an architecture argument, not a marketing one, and it is the number worth carrying into a cost model.

The Two Releases, Side by Side

Two details in that table trip people up:

• "Official release" and "GA" are not the same word, and DeepSeek chose them deliberately. The changelog describes Flash-0731 as an official release with the API in public beta, while the August 13 entry says the GA release of V4-Pro rolled out across app, web, and API. If your procurement process cares about the GA label specifically, Pro clears it and Flash does not yet.
• The parameter count depends on what you're counting. Both 284B and 304B circulate for Flash because the Hugging Face repo bundles a speculative decoding draft module on top of the 284B base. NVIDIA's NIM model page lists it as a 284B MoE with 1M context. Pick the base number and say what you mean by it.

Why This Matters?

Before the technical breakdown, it's worth translating a few terms that carry real weight in this release but rarely get explained in coverage of it.

• Mixture-of-Experts (MoE) means the model isn't one giant network processing every request in full. It's built from many smaller "expert" sub-networks, and each request only activates a fraction of them. This is why you'll see two different parameter counts for the same model: the total size (all experts combined) and the activated size (what actually runs for a given request). V4-Flash's 284B figure is its activated size, not its full footprint, which is exactly why it can run faster and cheaper than a dense model of the same total scale.
• A 1M-token context window is large enough to hold roughly 700,000 to 750,000 words in one request, several long novels, a mid-sized codebase, or a stack of legal contracts, all processed together rather than chunked and summarized piecemeal. That matters practically because chunking a document loses cross-references; a model that can hold the whole thing at once can answer questions that span the beginning and the end of a 400-page filing without losing the thread.
• KV cache is the memory a model has to keep in GPU memory to track everything it has already read in the current conversation or document. It grows with context length, and it's often the actual bottleneck on cost and hardware requirements at large context sizes, more so than raw compute. This is why the 10% KV cache figure matters as much as the 27% FLOPs figure: a model that needs a tenth of the memory to hold the same million tokens can run on meaningfully cheaper hardware, or serve far more concurrent users on the same hardware.
• GA (General Availability) is a specific commitment, not marketing language. It signals a vendor stands behind a release for production use, with the stability and support expectations that implies. A "public beta" API, by contrast, is explicitly still being tested and can change without the same guarantees, which is the practical reason the GA-versus-beta distinction covered throughout this piece actually matters for anyone building something meant to stay running.

With those terms in hand, the rest of this breakdown should read less like a spec sheet and more like what it actually is: two models built the same way, released two weeks apart, with real differences in what you can rely on from each.

Can You Use DeepSeek V4 for Free?

• V4-Flash-0731: yes, if you have the hardware. The weights and repository are MIT licensed, the permissive end of the license spectrum, allowing commercial use, modification, and redistribution with no copyleft obligation. The download is 167 GB across 48 safetensors shards, so "free" here means free of license fees, not free of a GPU bill.
• V4-Pro: a different proposition entirely. It's served through DeepSeek's app, web interface, and API. First-party weights for the 0813 checkpoint were not something we could confirm on DeepSeek's own Hugging Face org while writing this. Third-party posts claim a DeepSeek-V4-Pro-0813 repo appeared on August 13. Treat that as unverified until you see the repo yourself.

The Pricing Change Is the Part With a Deadline

The GA post carries new pricing effective 16:00 UTC on August 16, 2026, and the structure changed rather than just the numbers:

• V4 moved to peak and off-peak rates, with off-peak set at half of peak.
• The changelog confirms this directly: with the official release of the V4 family, pricing moves to a peak and off-peak model.

That has a practical consequence for anyone running batch work. Evaluation sweeps, backfills, document pipelines, and nightly agent runs are exactly the jobs that can be shifted into an off-peak window, halving their token cost with no code change beyond a scheduler. Interactive traffic cannot move, so the saving is entirely a question of what share of your calls are latency-insensitive.

Specific per-million-token figures are circulating from aggregator pages rather than DeepSeek's own pricing docs. We are not quoting them here. Read the rate off DeepSeek's pricing page before you build a forecast on it, since the peak and off-peak split means a single blended number is misleading by design.

What Is DeepSeek V4-Pro Used For?

V4-Pro is positioned for agent work, coding, and long-context reasoning. The GA post leads with major agent upgrades, and the front page of deepseek.com describes the official V4-Pro as having significantly enhanced agent capabilities with support for the Responses API and Codex.

The Responses API detail is the one developers should notice. Native support for that request format, plus Codex adaptation, means V4-Pro can drop into tooling written against that interface without a translation layer. DeepSeek's own X announcement on July 31 made the same point about Flash: native Responses API format support and Codex adaptation, with the calling method otherwise unchanged.

Thinking effort is now explicit, with three levels for both models:

• Low
• High
• Max

This is a knob, not a mode switch, and it interacts with the pricing change directly: max effort on peak-rate tokens is the most expensive combination available in the family.

How the Two Models Divide the Work

The interesting claim in DeepSeek's own Flash-0731 model card is that Flash outperforms V4-Pro Preview on the listed benchmarks despite far fewer activated parameters. Note the comparison target: Pro Preview, not the GA Pro-0813 that shipped two weeks later. Flash beating the preview flagship is not Flash beating the shipped flagship.

A reasonable division of labor, based on what each release actually claims:

• High-volume agent loops and coding assistance, where token throughput dominates: V4-Flash, self-hosted if you have the GPUs, API if you do not.
• Long-context analysis over large document sets, where the 27% FLOPs and 10% KV cache figures pay off: V4-Pro.
• Anything with a data residency requirement that rules out a hosted endpoint: V4-Flash, since MIT weights are the only option in the family you can run entirely inside your own perimeter.
• Anything needing a GA-labeled dependency: V4-Pro, since Flash's API is still public beta.

Treat the Benchmark Numbers as Vendor-Reported

This is where most coverage of the V4 GA overreaches, so it is worth being precise about what the evidence actually is.

DeepSeek's benchmark tables are first-party and self-run:

• The model card states that code-agent tasks were evaluated using the minimal mode of DeepSeek Harness at max reasoning effort.
• The harness itself is described as still to be released.
• Two of the referenced test sets, DSBench-FullStack and DSBench-Hard, are internal.

None of that makes the numbers false. It makes them unreplicable right now, which is a different problem. Until the harness ships and an independent lab runs it, any "V4-Pro beats X on Y" line in a blog post is repeating DeepSeek's own measurement with the attribution stripped off. If you are writing a business case, cite the architecture figures from the arXiv paper, which are checkable, and hold the agent benchmark claims until third-party runs exist.

One more thing to keep straight: there was genuine disagreement in late-July coverage about whether the -0731 weights were on Hugging Face the same day the API went to beta, with some posts on July 31 reporting that the downloadable version was still the April preview. The repo exists now and it is MIT licensed. The exact upload hour is not settled by first-party evidence, so do not build a timeline claim on it.

What About the Region-by-Region Rollout?

Second-hand roundups describe a continuing region-by-region grayscale rollout of V4 GA. We could not find that in any DeepSeek source. The GA post and the changelog both describe the rollout as done across app, web, and API on August 13, with no staged regional language.

If you are seeing V4-Pro behave differently on two accounts, the likelier explanation is the pricing cutover on August 16 or the thinking effort default, not a hidden regional gate. Check the model string your client is actually sending first: deepseek-v4-pro resolves to the 0813 checkpoint, and code written against the preview may still be pinned to an older string.

Will There Be a DeepSeek V5?

There is no date in any DeepSeek first-party source. Analyst posts predict late August or early September 2026 based on the V4 cadence and price competition among Chinese labs, which is a forecast, not an announcement.

The signal worth watching instead is the harness. DeepSeek said DeepSeek Harness is to be released, and a public harness is what turns the current self-reported agent scores into something anyone can reproduce. That release would tell you more about the V4 family's real standing than a V5 rumor does.

How This Changes the Practical Decision

If you are choosing a model this month, the V4 releases shift three things:

1. The open-weight ceiling moved. A 284B MoE with a 1M-token context under MIT is a materially different offer from what was downloadable in April, and it lands in the same window as deepseek and chatgpt comparisons that predate it and are now out of date on specifics.
2. Batch economics changed shape. Peak and off-peak at a 2:1 ratio rewards scheduling, a rare case of a pricing change you can capture with a cron entry.
3. The GA label and the beta label now sit on different tiers of one family, so "we are on DeepSeek V4" is no longer a precise statement about anything. Name the checkpoint.

For creative work, none of this replaces a purpose-built visual stack. V4 is a text and agent family, so image and video generation still belong to dedicated models, and if that is your use case an AI image generator is the relevant surface rather than a 1.6T-parameter language model. Where V4 does earn a place in a creative pipeline is the orchestration layer: scripting, prompt generation, metadata, and the kind of content automation that runs unattended and can be pushed into off-peak hours.