DeepSeek V4 is the open-weight flagship family that ended the long wait for DeepSeek R2 — and, on the evidence of its benchmarks, pricing and adoption, the most consequential large language model release of 2026 so far. The series spans two Mixture-of-Experts models: DeepSeek V4 Pro, which reached general availability on 13 August 2026, and DeepSeek V4 Flash, finalised on 31 July 2026. Both ship MIT-licensed open weights, a one-million-token context window and output pricing that currently sits below one dollar per million tokens.
The timing could hardly be fresher. DeepSeek published its general-availability announcement for V4-Pro-0813 on 13 August 2026 — the day before this guide went live — highlighting agent capabilities and what the company calls “strong production gains”. The same release introduced flexible reasoning effort levels, folding the machine learning ambitions once earmarked for a standalone R2 directly into DeepSeek V4.
We have tracked this family on our AI models and tools hub since the April preview, including our May report on the DeepSeek V4 API price cut for coding agents. That price story now has a sequel: from 16 August 2026 DeepSeek switches to an unusual peak/off-peak billing model, and the peak rates are a substantial rise on today’s prices.
This DeepSeek V4 complete guide walks through the whole picture: the two builds and their release timeline, the R2 saga that shaped them, the architecture behind the million-token context, verified benchmark scores, current and future pricing, the licence position, developer integration, market share, and the government restrictions any adopter should weigh. Every figure comes from the sources listed in the References section.
Table of contents
- What Is DeepSeek V4? Series Overview at a Glance
- The Road to V4: Why DeepSeek R2 Never Shipped
- DeepSeek V4 Architecture: Inside the Technical Report
- Million-Token Context and Efficiency Gains
- DeepSeek V4 Benchmarks: Pro vs Flash in Numbers
- Reasoning Effort Levels Explained
- DeepSeek V4 Pricing: Rates, Caching and the Peak/Off-Peak Shift
- Licence and Open-Weight Status
- DeepSeek V4 for Developers and Agent Builders
- Getting Started: Three Routes into the Models
- Market Position and Adoption in 2026
- How V4 Stacks Up in the Open-Weight Field
- Government Restrictions and Controversies
- Hardware, Funding and the Road Ahead
- How to Choose Between DeepSeek V4 Pro and Flash
- DeepSeek V4 FAQ
- References
What Is DeepSeek V4? Series Overview at a Glance
DeepSeek’s official API price list carries exactly two production models as of August 2026: deepseek-v4-pro, which points to DeepSeek-V4-Pro-0813, and deepseek-v4-flash, which points to DeepSeek-V4-Flash-0731. There is no separate reasoning model and no R2 — the DeepSeek V4 series is the entire current line.
The family launched in preview on 24 April 2026 as two open-weight models released simultaneously on Hugging Face, the API and chat.deepseek.com. The preview label mattered: DeepSeek spent the following three and a half months hardening both builds before declaring them final.
General availability is more than a label change. It fixes the checkpoints that enterprises can pin, ends the behaviour drift of the preview months, and signals that DeepSeek will stand behind these builds in production. For teams that held off during the preview period, the 0813 and 0731 builds are the versions to evaluate.
| Specification | DeepSeek V4 Pro (0813) | DeepSeek V4 Flash (0731) |
|---|---|---|
| General availability | 13 August 2026 | 31 July 2026 |
| Total parameters (technical report) | 1.6 trillion | 284 billion |
| Activated parameters | 49 billion | 13 billion |
| GA checkpoint size (Hugging Face) | 1.7T (DSpark attached) | 304B (DSpark attached) |
| Context window | 1 million tokens | 1 million tokens |
| Maximum output | 384K tokens | 384K tokens |
| Weights licence | MIT License | MIT License |
| API model id | deepseek-v4-pro | deepseek-v4-flash |
DeepSeek V4 Pro 0813: The General-Availability Flagship
DeepSeek V4 Pro is the heavyweight of the pair: a Mixture-of-Experts design with 1.6 trillion total parameters, of which only 49 billion activate per token. The 0813 GA build is the version DeepSeek now recommends for production, with the launch notes emphasising agent capabilities and strong production gains over the April preview.
The GA release also made DeepSeek V4 Pro natively compatible with the OpenAI Responses API, with specific optimisation for Codex integration — a clear signal about which workloads DeepSeek expects to win.
DeepSeek V4 Flash 0731: The Efficient Workhorse
DeepSeek V4 Flash carries 284 billion total parameters with just 13 billion activated. Despite that far smaller activated parameter count, DeepSeek describes the 0731 build as outperforming its own preview version and as “broadly competitive with the strongest proprietary models available”.
For most day-to-day workloads — summarisation, extraction, routine coding, high-volume agent steps — Flash is the sensible default, at a fraction of the price of Pro. We compare the two head-to-head later in this guide.
From April Preview to August GA
The DeepSeek V4 timeline runs in four steps. The preview pair arrived on 24 April 2026. On 4 July 2026, separate DSpark speculative-decoding variants appeared on Hugging Face. DeepSeek-V4-Flash-0731 became the official post-preview Flash build on 31 July. Finally, DeepSeek-V4-Pro-0813 reached general availability on 13 August 2026, completing the series.
The Road to V4: Why DeepSeek R2 Never Shipped
You cannot understand DeepSeek V4 without the ghost that haunts it. Through 2025, the AI world expected DeepSeek R2 — the successor to the R1 reasoning model that had rattled markets. R2 never arrived, and the reasons explain much of what V4 became.
The R2 That Never Arrived
As of mid-2026, DeepSeek R2 has no technical report, no benchmarks, no pricing and no launch date, and it does not appear anywhere in DeepSeek’s official API model list. It is, in every practical sense, a model that never shipped. Analysts who kept waiting for an “R2 moment” watched DeepSeek take a different road entirely.
The absence mattered because expectations were enormous. R1 was the release that made DeepSeek a global name, and the assumption through 2025 was that a successor would follow the same playbook: a dedicated reasoning model, trained separately and sold separately. Every quarter that passed without it fed speculation about training troubles — speculation that later reporting substantially confirmed.
Huawei Ascend Setbacks and the Switch Back to Nvidia
Reporting by Tom’s Hardware traced the delay to hardware. After Chinese authorities encouraged DeepSeek to adopt Huawei silicon, R2 training reportedly stalled on Huawei Ascend 910C processors following multiple failures. DeepSeek ultimately reverted to Nvidia hardware for training, while continuing to adapt Ascend chips for inference workloads.
That split — train on Nvidia, serve on Ascend — quietly became the template for the DeepSeek V4 era, and it resurfaces later in this guide when we look at Huawei’s Ascend 950 supernode support.
There is a broader lesson for anyone planning multi-year AI roadmaps: model schedules are hostage to silicon. A lab with world-class researchers still slipped a full product generation when its training hardware underperformed. Buyers evaluating vendor promises about future models should discount them accordingly.
How Reasoning Folded into DeepSeek V4
Rather than shipping a standalone reasoning model, DeepSeek built the capability into V4 itself as operating modes: non-reasoning, reasoning, and a “Max” extended-reasoning mode. The 0813 GA release refined this further into flexible reasoning effort levels for both builds. Strategically, DeepSeek V4 is R2 — dissolved into a general-purpose model rather than sold as a separate product.
DeepSeek V4 Architecture: Inside the Technical Report
The official technical report — “DeepSeek-V4: Towards Highly Efficient Million-Token Context Intelligence”, published as arXiv:2606.19348 — is unusually explicit about how the efficiency was achieved. Four ingredients stand out.
A Sparse Mixture of Experts at 1.6-Trillion Scale
Both DeepSeek V4 models are Mixture-of-Experts systems that activate only a small slice of their weights per token: 49 billion of 1.6 trillion for Pro, 13 billion of 284 billion for Flash. That roughly 3% activation ratio on Pro is what lets a 1.6T-parameter model run at the serving cost of a much smaller dense network.
The economics follow directly. In a dense model, every parameter participates in every token, so serving cost scales with total size. In a Mixture-of-Experts design, a router selects a handful of specialist expert networks per token, so cost scales with the activated slice instead. Total parameters buy breadth of knowledge; activated parameters set the per-token bill. DeepSeek V4 pushes that ratio further than any open-weight family before it.
Hybrid Attention: CSA and HCA
For long-context efficiency, DeepSeek V4 uses a hybrid attention architecture combining Compressed Sparse Attention (CSA) with Heavily Compressed Attention (HCA). The pairing is the key enabler of the million-token window: it attacks both the compute and the memory cost of attending over extremely long sequences.
Hyper-Connections and the Muon Optimiser
The report describes two further departures from convention. Standard residual connections are replaced with Manifold-Constrained Hyper-Connections (mHC), and training used the Muon optimiser — notable because DeepSeek ran Muon at 1.6-trillion-parameter scale, well beyond where it had been proven before.
Why DeepSeek V4 Has Two Parameter Counts
Careful readers will notice a wrinkle: the technical report says 1.6T and 284B, but the GA checkpoints on Hugging Face list DeepSeek-V4-Pro-0813 at 1.7 trillion parameters and DeepSeek-V4-Flash-0731 at 304 billion. Neither number is wrong.
DeepSeek also published separate DSpark variants on 4 July 2026 — DeepSeek-V4-Pro-DSpark at 1.7T and DeepSeek-V4-Flash-DSpark at 165B — for teams that want the speculative-decoding configuration as a standalone artefact.
Speculative decoding, for readers new to the term, is a serving trick rather than a quality change: a small draft model proposes several tokens ahead, and the main model verifies them in one pass, accepting the correct ones. When the drafts are mostly right, throughput jumps and latency drops with identical output quality — which is why DeepSeek ships DSpark both attached to the GA checkpoints and as standalone variants.
Million-Token Context and Efficiency Gains
If one theme defines this generation, it is efficiency at extreme context length. DeepSeek V4 was engineered so that a million-token window is not a marketing checkbox but something you can actually afford to serve.
One Million Tokens in Every DeepSeek V4 Model
Both models support a context length of one million tokens, and the API exposes that full window along with a maximum output of 384K tokens for deepseek-v4-pro and deepseek-v4-flash alike. Whole codebases, long legal bundles and multi-day agent transcripts fit into a single DeepSeek V4 call without chunking gymnastics.
What does that buy in practice? A million tokens comfortably holds a mid-sized monorepo, an entire discovery bundle in litigation, or weeks of accumulated agent memory. The deeper win is architectural: retrieval pipelines built to work around small windows can be simplified or removed, because the model can simply be shown everything relevant at once.
27% of the FLOPs, 10% of the KV Cache
The technical report’s headline efficiency claim: at one-million-token context, DeepSeek V4 Pro requires only 27% of the single-token inference FLOPs and just 10% of the KV cache of DeepSeek-V3.2. The chart below visualises those reported ratios against the V3.2 baseline.
The KV-cache figure matters most for operators: memory, not raw compute, is usually what caps long-context concurrency. Cutting the cache to a tenth changes what a single serving node can hold.
A 32-Trillion-Token Training Corpus
The efficiency did not come at the expense of training scale. The V4 models were pre-trained on more than 32 trillion tokens of curated data — versus 14.8 trillion for DeepSeek-V3, according to that model’s own report. In other words, DeepSeek more than doubled the corpus between generations while driving serving costs down.
DeepSeek V4 Benchmarks: Pro vs Flash in Numbers
Benchmarks are where the “best open-weight AI of 2026” claim earns its keep. The scores below come from the official model cards for the two GA builds; treat them as vendor-reported until independent replications accumulate.
DeepSeek V4 Pro Benchmark Scores
On its model card, DeepSeek V4 Pro 0813 posts 60.0 on Humanity’s Last Exam with tools and 87.9 on Terminal Bench 2.1. The coding and agent picture is equally strong: 61.5 on NL2Repo, 83.3 on Cybergym and 62.7 on DeepSWE, with 74.1 on Toolathlon-Verified, 71.1 on DSBench-FullStack and 67.2 on DSBench-Hard.
DeepSeek V4 Flash Benchmark Scores
DeepSeek V4 Flash 0731 lands closer to its bigger sibling than its 13B activated parameters suggest: 82.7 on Terminal Bench 2.1, 54.2 on NL2Repo, 76.7 on Cybergym, 54.4 on DeepSWE, 70.3 on Toolathlon-Verified, 68.7 on DSBench-FullStack and 59.6 on DSBench-Hard.
| Benchmark | DeepSeek V4 Pro 0813 | DeepSeek V4 Flash 0731 |
|---|---|---|
| Humanity’s Last Exam (with tools) | 60.0 | not published |
| Terminal Bench 2.1 | 87.9 | 82.7 |
| NL2Repo | 61.5 | 54.2 |
| Cybergym | 83.3 | 76.7 |
| DeepSWE | 62.7 | 54.4 |
| Toolathlon-Verified | 74.1 | 70.3 |
| DSBench-FullStack | 71.1 | 68.7 |
| DSBench-Hard | 67.2 | 59.6 |
On the two agentic headline tests, the gap between the builds is narrow — as the chart of model-card scores below shows.
Reading the Numbers Fairly
Two caveats keep this honest. First, all figures above are self-reported model-card scores; independent leaderboards will take weeks to catch up with the GA builds. Second, Flash’s Humanity’s Last Exam score is simply not published, so the table leaves it blank rather than guessing. Even so, DeepSeek’s own framing — Flash as “broadly competitive with the strongest proprietary models available” — is a bold public claim for an open-weight 13B-activation model.
It is also worth noticing which benchmarks DeepSeek chose to publish. Terminal Bench, Toolathlon, NL2Repo, DeepSWE and DSBench all measure a model doing things — operating terminals, calling tools, editing repositories — rather than answering exam questions. That selection mirrors how buyers actually deploy models in 2026, and it is exactly where DeepSeek V4 wants to be judged.
Reasoning Effort Levels Explained
The most distinctive product decision in this generation is how reasoning is exposed. Instead of choosing between a chat model and a reasoning model, you choose how hard one model thinks.
Low, High and Max: Three Effort Settings
The 0813 GA release introduced flexible reasoning effort for both builds: low for simple tasks, high for daily agent workflows, and max for complex tasks. This sits on top of the operating modes the family already had — non-reasoning, reasoning, and the extended-reasoning “Max” mode. Billing stays token-based, so effort translates directly into cost you control per request.
The dial matters because reasoning tokens are billed like any others. An agent that thinks at maximum effort on every trivial step burns budget and adds latency for no quality gain; one that never thinks hard fails the steps that matter. Making effort a per-request parameter moves that trade-off out of model selection and into ordinary engineering.
How DeepSeek V4 Compares with Kimi K3’s Approach
DeepSeek is not alone here: Kimi K3 ships a directly comparable reasoning-effort system, which we analysed when it launched in July. The convergence is telling — two of China’s leading labs independently concluded that reasoning should be a dial, not a separate model. For agent builders, DeepSeek V4 turns “how much thinking is this step worth?” into an explicit engineering parameter.
DeepSeek V4 Pricing: Rates, Caching and the Peak/Off-Peak Shift
Pricing has always been DeepSeek’s sharpest weapon, and the DeepSeek V4 price list remains startling reading next to Western flagships. But the story is changing on 16 August 2026, and anyone budgeting around today’s rates needs to read this section carefully.
Current DeepSeek V4 API Pricing
Until 16 August, deepseek-v4-pro costs $0.435 per million input tokens on a cache miss, $0.003625 per million on a cache hit, and $0.87 per million output tokens. For deepseek-v4-flash the rates are $0.14, $0.0028 and $0.28 respectively. Third-party pricing trackers independently corroborate the Pro figures, so these are not list prices that nobody pays.
| Model (per 1M tokens) | Input, cache miss | Input, cache hit | Output |
|---|---|---|---|
| DeepSeek V4 Pro | $0.435 | $0.003625 | $0.87 |
| DeepSeek V4 Flash | $0.14 | $0.0028 | $0.28 |
The cache-hit price deserves a highlight: at $0.003625 per million tokens on Pro, repeated context — system prompts, tool schemas, shared codebases — becomes essentially free. Agent loops that resend large contexts benefit enormously.
The Peak and Off-Peak Experiment from 16 August
From 16 August 2026 at 16:00 UTC, DeepSeek switches to peak/off-peak billing — a first among major AI providers. Peak hours are 01:00–04:00 and 06:00–10:00 UTC, and off-peak rates run at half the peak rates.
| Model and window (per 1M tokens) | Input, cache miss | Output |
|---|---|---|
| V4 Pro — peak | $1.32 | $3.96 |
| V4 Pro — off-peak | $0.66 | $1.98 |
| V4 Flash — peak | $0.44 | $1.32 |
| V4 Flash — off-peak | $0.22 | $0.66 |
Make no mistake: this is a rise. Pro’s peak output rate of $3.96 is roughly four and a half times the current $0.87, and even the off-peak $1.98 is more than double today’s price. The May price cut we reported earlier has, in effect, been unwound for peak-hour traffic.
Why would an AI provider borrow electricity-market pricing? Because GPU clusters, like power stations, are capacity businesses: demand peaks force overprovisioning that then sits idle off-peak. Charging half price outside the busy windows shifts elastic workloads — batch jobs, evaluations, overnight agent runs — into the troughs, flattening the load curve. If the experiment works, expect other providers to copy it.
Budgeting for the New Rates
The practical playbook writes itself. Batch and scheduled workloads should move outside 01:00–04:00 and 06:00–10:00 UTC to capture the half-price window. Interactive products serving users during UTC peak hours should lean harder on Flash and on prompt caching. And any cost model built on the pre-16-August rates needs revisiting this week — a deployment that was priced in July is mispriced in September.
Licence and Open-Weight Status
For many adopters, the licence is the whole reason DeepSeek V4 is on the shortlist. This is one of the few frontier-class families you can download, modify and ship commercially without negotiating anything.
MIT-Licensed Weights on Hugging Face
The V4 weights are released on Hugging Face under the MIT License — one of the most permissive licences in software. There is no usage-tier cut-off, no separate commercial agreement and no copyleft obligation attached to the weights. That applies to both GA checkpoints and, per the deepseek-ai organisation page, the DSpark variants as well.
Among open licences, MIT sits at the maximally permissive end: attribution is essentially the only obligation. There is no requirement to publish derivatives, no field-of-use restriction and no revenue threshold that converts the licence into a commercial negotiation — conditions that several rival open-weight families do attach. For legal teams, that simplicity is often the fastest part of the whole approval process.
What Open Weights Mean for DeepSeek V4 Deployments
Open weights are not the same as open source — the 32-trillion-token training corpus is not published — but for deployment purposes the distinction rarely bites. You can self-host DeepSeek V4 on your own infrastructure, fine-tune it on private data, and keep every token inside your network boundary. As the restrictions section below shows, that self-hosting option is also the cleanest answer to the data-governance concerns that surround DeepSeek’s hosted service.
DeepSeek V4 for Developers and Agent Builders
The GA announcement’s emphasis on agent capabilities was not incidental. Everything about this release — benchmarks chosen, API surface, effort levels — targets teams building autonomous coding and tool-use systems.
OpenAI-Compatible Responses API and Codex
DeepSeek V4 Pro 0813 ships with native OpenAI Responses API compatibility, specifically optimised for Codex integration. Practically, that means agent stacks written against OpenAI’s interface can point at deepseek-v4-pro with minimal rework — a deliberate lowering of the switching cost from proprietary incumbents.
Building Agent Workflows on DeepSeek V4
The ingredients stack neatly: Terminal Bench 2.1 at 87.9 and Toolathlon-Verified at 74.1 for tool-heavy work, a million-token window for long-horizon task state, near-free cache hits for resent context, and reasoning effort as a per-step dial. Teams we work with through our intelligent automation practice typically route routine steps to Flash on low effort and escalate to Pro on max effort only when a step genuinely demands it.
Self-Hosting and Infrastructure Choices
Self-hosters should start from the parameter table: Pro’s 1.6T total weights demand serious multi-node infrastructure even though only 49B activate per token, while Flash at 284B total is within reach of a far wider range of clusters. The DSpark speculative-decoding variants published on 4 July exist precisely to squeeze more throughput from the same hardware. If you are weighing managed API against self-hosted weights, our ML model development team runs exactly this evaluation for clients.
Budget for the operational reality, though. Serving a long-context Mixture-of-Experts model well means expert-parallel inference, KV-cache management and careful batching — a different discipline from running a small dense model on a single card. Many teams land on a hybrid: the hosted API for spiky interactive traffic, self-hosted weights for steady bulk workloads where the economics favour owned hardware.
Getting Started: Three Routes into the Models
Everything above describes the models; this section is about your first week with them.
Chat, API or Weights
There are three ways in, and they suit different stages of adoption. The quickest look is chat.deepseek.com, where the family has been available since the April preview. The API is the production route: point your client at the deepseek-v4-pro or deepseek-v4-flash model ids and you are on the GA builds automatically. The third route — downloading the MIT-licensed checkpoints from Hugging Face — is for teams with the infrastructure to serve DeepSeek V4 themselves.
Fitting It into an Existing Stack
If your stack already speaks OpenAI’s interfaces, the Responses API compatibility discussed earlier makes the swap mostly a configuration change: change the base URL and model id, keep the calling code. The details worth testing are the ones compatibility layers historically fumble — streaming behaviour, tool-call formats and error semantics — plus the effort parameter, which has no direct OpenAI equivalent.
A Sensible First Evaluation
Resist the urge to start with a leaderboard replica. Take ten real tasks from your own backlog — the tickets, documents and queries your team actually handles — and run them through Flash at low and high effort, then through Pro at max. Score the outputs blind. This costs a few dollars at current rates and tells you more than any public benchmark about how DeepSeek V4 behaves on your distribution.
Common Early Mistakes
Three mistakes recur in evaluations we see. First, testing only at default effort, which understates what the max mode can do on hard problems. Second, ignoring the prompt cache: at $0.003625 per million cached input tokens on Pro, a well-structured system prompt is nearly free, while a badly structured one silently multiplies costs. Third, benchmarking in August against prices that change on the 16th — model the peak/off-peak rates, not the ones being retired.
Market Position and Adoption in 2026
Capability claims are easy; token volume is harder to fake. On neutral routing platforms, DeepSeek’s models were already a dominant force months before the GA builds landed.
DeepSeek V4 on OpenRouter: The Token-Share Story
In the week of 24 February 2026, Chinese-built models accounted for 61% of token volume among OpenRouter’s top-10 models. DeepSeek alone held 17.6% weekly platform share — more than Google (12.5%) and OpenAI (8.4%) combined — occupying ranks 1, 5 and 8 and processing 1.08 trillion tokens in a single day. That snapshot predates the V4 preview, which makes it more striking: the audience DeepSeek V4 inherited was already the platform’s largest.
Token share is a meaningful metric precisely because it is behavioural. Developers on routing platforms pay per call and can switch models by editing one string, so sustained share reflects thousands of independent cost-quality decisions rather than any single procurement deal. It is the closest thing the model market has to a revealed-preference poll.
The Efficiency-First Race
Forbes framed the April launch as proof that “the next AI race is about efficiency”, and the thesis has aged well. While Western labs competed on raw scale, DeepSeek V4 competed on cost per useful token — 27% of the FLOPs, 10% of the KV cache, sub-dollar output pricing. The peak/off-peak experiment extends the same logic from model design into demand shaping.
How V4 Stacks Up in the Open-Weight Field
Readers of our hub will know 2026’s open-weight field is crowded: we cover Qwen, Kimi, GLM and Mistral releases alongside the proprietary flagships. Placing this family in that field without cherry-picking is difficult — vendors report different benchmarks — but some differentiators are objective.
The Differentiators That Matter
Three things are simply facts rather than marketing. The MIT licence is more permissive than most rival open-weight terms. The one-million-token context window with 384K output sits at the top of what any open-weight family currently offers. And the pricing table above — under a dollar per million output tokens today, with a cache-hit input rate measured in tenths of a cent — is the sharpest economic pitch among frontier-class models. Where benchmarks are contested, economics are not.
Effort Levels as an Industry Pattern
The reasoning-effort convergence we noted with Kimi K3 looks increasingly like the industry’s settled answer to the reasoning-model question. Separate reasoning products create routing headaches; a dial on one model does not. That DeepSeek folded R2 into the V4 line rather than shipping it standalone is the strongest evidence of where the design consensus has landed.
Where Competitors Still Lead
Honesty requires the converse list. Rival families publish strengths of their own on benchmarks DeepSeek does not report, and the V4 line ships no small on-device variant — Flash’s 284 billion total parameters are the family’s floor. If a workload needs a pocket-sized model or capabilities the model cards above do not cover, the hub’s other guides map the alternatives.
Government Restrictions and Controversies
No honest guide to this family can skip this section. The model family’s technical merits coexist with serious geopolitical friction, and for some organisations that friction is disqualifying.
The US NDAA Removal Order
The US National Defense Authorization Act for Fiscal Year 2026 ordered DeepSeek AI removed from Department of Defense and Intelligence Community systems, allowing only limited exceptions for counterterrorism and counterintelligence research. US federal contractors in the defence orbit should treat that as a bright line.
Australia’s Directive
Australia moved earlier and broader at government level: the Department of Home Affairs issued a directive preventing the use or installation of DeepSeek products across government systems. Other jurisdictions have taken varying positions, so multinationals need a per-country view before standardising on the family.
Weighing DeepSeek V4 Risk in Your Stack
The pragmatic reading: the bans above target government systems and DeepSeek’s products as hosted services. Because DeepSeek V4 weights are MIT-licensed, a self-hosted deployment inside your own network answers the data-flow half of the concern — no tokens ever reach DeepSeek’s servers. It does not answer procurement policy, which is a legal and reputational judgement each organisation must make for itself.
A sensible governance posture treats model choice like any other supplier decision: document where inference runs, what data leaves your boundary, and which jurisdictions’ rules apply to each workload. Teams that write this down per use case usually find the question far less binary than headlines suggest — some workloads clear easily, others never will.
Hardware, Funding and the Road Ahead
The V4 generation is as much an industrial story as a research one, and two threads point at where DeepSeek goes next.
Huawei Ascend 950 Supernodes and Co-Design
Huawei has said its Ascend 950-based supernode clusters fully support the V4 series — positioning DeepSeek V4 within a deliberate China hardware co-design strategy. Combined with the train-on-Nvidia, serve-on-Ascend split that emerged from the R2 saga, the direction is clear: DeepSeek is engineering its models to be efficient on domestic silicon, insulating its serving stack from export controls.
IPO Signals and Funding
The corporate backdrop is heating up. In April 2026, investors discussed a $300 million funding round targeting a roughly $10 billion valuation. By July, Bloomberg and the Financial Times were reporting that DeepSeek is preparing for an IPO as soon as 2027. A listed DeepSeek would face new disclosure pressures — which, for enterprise adopters of DeepSeek V4, would arguably be a transparency upgrade.
How to Choose Between DeepSeek V4 Pro and Flash
Most evaluation cycles end at the same question: which build, for which workload? The published numbers make the trade-off unusually easy to reason about.
When DeepSeek V4 Pro Is Worth It
Choose DeepSeek V4 Pro when the task sits at the frontier of what agents can do: multi-step repository work (NL2Repo 61.5 vs 54.2), hard security tasks (Cybergym 83.3 vs 76.7), or DSBench-Hard-style problems where the 67.2-to-59.6 gap compounds across a pipeline. From 16 August, running it off-peak at $1.98 output softens the premium considerably.
When Flash Is Enough
DeepSeek V4 Flash at 82.7 on Terminal Bench 2.1 is within striking distance of Pro at a fraction of the price — $0.28 output today, $0.66 off-peak after the change. For high-volume agent steps, extraction, summarisation and routine coding, Flash is the rational default. DeepSeek’s own claim that it is broadly competitive with the strongest proprietary models is aggressive, but the model-card deltas back the spirit of it.
A Simple Decision Framework
Start every workload on Flash at low effort. Escalate individual steps to high effort when tool use gets complex. Reserve Pro on max effort for the steps where benchmark gaps are widest — hard repository edits and security-sensitive tasks — and schedule batch jobs off-peak. Teams that follow this pattern get most of Pro’s quality at something close to Flash’s cost.
DeepSeek V4 FAQ
Is DeepSeek V4 free to use?
The weights are free: both GA checkpoints are on Hugging Face under the MIT License, so you can download and self-host DeepSeek V4 at no licence cost. The hosted API is paid per token — currently $0.87 per million output tokens for Pro and $0.28 for Flash — and the family is also available through chat.deepseek.com.
Did DeepSeek ever release R2?
No. As of mid-2026, DeepSeek R2 has no technical report, benchmarks, pricing or launch date, and it is absent from the official API model list. Its reasoning ambitions shipped instead as modes and effort levels inside the V4 family, after R2 training reportedly stalled on Huawei Ascend 910C hardware.
What context window does DeepSeek V4 support?
Both models support one million tokens of context, with the API allowing up to 384K output tokens. Per the technical report, DeepSeek V4 Pro delivers that window using only 27% of the inference FLOPs and 10% of the KV cache of DeepSeek-V3.2.
Which is better: DeepSeek V4 Pro or DeepSeek V4 Flash?
Pro wins on every published shared benchmark — Terminal Bench 2.1 87.9 vs 82.7, Cybergym 83.3 vs 76.7, DeepSWE 62.7 vs 54.4 — but Flash costs a fraction as much per token. Use Flash as the default and reserve Pro for the hardest agentic steps.
Is DeepSeek V4 banned anywhere?
In specific contexts, yes. The US NDAA for Fiscal Year 2026 ordered DeepSeek removed from Department of Defense and Intelligence Community systems, and Australia’s Department of Home Affairs blocked DeepSeek products across government systems. Self-hosting the MIT-licensed weights addresses data-flow concerns but not procurement policy.
When does peak/off-peak billing start?
On 16 August 2026 at 16:00 UTC. Peak hours run 01:00–04:00 and 06:00–10:00 UTC, with off-peak rates at half the peak rates. Peak output pricing rises to $3.96 per million tokens for Pro and $1.32 for Flash — a substantial increase on current rates, so schedule batch workloads off-peak.
References
DeepSeek API News — V4-Pro GA Announcement (13 August 2026)
DeepSeek API Pricing (Official)
DeepSeek-V4-Pro-0813 Model Card — Hugging Face
DeepSeek-V4-Flash-0731 Model Card — Hugging Face
DeepSeek-V4: Towards Highly Efficient Million-Token Context Intelligence (arXiv:2606.19348)
deepseek-ai Organisation — Hugging Face
Forbes — DeepSeek V4 Shows That The Next AI Race Is About Efficiency
Tom’s Hardware — DeepSeek R2 Training Moved Back to Nvidia Hardware
Dataconomy — Chinese AI Models Hit 61% Market Share on OpenRouter
More AI coverage: explore Progressive Robot's AI Models, Tools & Releases hub — hands-on reviews, setup guides and benchmarks in one place.
