DeepSeek model timeline
DeepSeek models: which version to use in 2026, and what's coming next
In a single year, DeepSeek went from R1 to a whole family of models. The latest confirmed release is DeepSeek V4 (V4-Pro and V4-Flash), launched on April 24, 2026, with a one-million-token context window and an MIT license. This page sorts it out: what actually exists, what's only a rumor (V5, R2), which version to choose for your needs, what it costs, and what you need to know about access and data privacy.
Every version
From R1 to V4: the DeepSeek family, made clear
DeepSeek ships fast. Here are the models that are actually available, newest to oldest, with what sets each apart and which one replaced which.
What to use today
V4-Pro for power, V4-Flash for everyday use
If you're starting now, these are the only two versions worth remembering. V4-Pro targets the heaviest coding and reasoning tasks. V4-Flash covers 90% of use cases at a fraction of the cost.
Both share the 1M-token context and the MIT license.
What's coming next
V5 and R2: rumors, no official date
No public roadmap confirms a DeepSeek V5, and R2 has been denied by the company several times. We track both here, keeping confirmed facts clearly separate from unverified rumors.
Updated as official announcements come in.
April 2026 - current version
DeepSeek V4-Pro
The flagship model: 1.6 trillion total parameters, 49 billion active per token. 80.6% on SWE-bench Verified, among the best coding scores in open source.
Ideal for autonomous code editing, long-form reasoning, and analyzing very large documents (1M-token context).
April 2026 - current version
DeepSeek V4-Flash
The fast, economical version: 284 billion parameters, 13 billion active. Same 1M context window as V4-Pro, at $0.14 per million input tokens.
It's the default chat model and the migration target for the old API identifiers as of July 24, 2026.
December 2025
DeepSeek V3.2
The first version to build reasoning directly into tool use. Simulated gold medals at the IMO and IOI. The technical base that V4 was built on.
Superseded by V4 for most use cases, but it laid down the efficient-attention architecture.
August 2025
DeepSeek V3.1
A hybrid model that merged V3 and R1 into a single system: 671 billion parameters, 37 billion active, 128K context. The step that made DeepSeek usable day to day.
A solid trade-off at the time, now outpaced by the V4 family on context and code.
January 2026
DeepSeek R1
The reasoning model that put DeepSeek on the map worldwide and sent Nvidia's stock tumbling. Visible chains of thought, ultra-low-cost training.
Many expect a successor, R2, but it remains unconfirmed (see the dedicated page).
The questions everyone's asking
DeepSeek in 2026: the straight answers
The most-searched questions on Google about DeepSeek models, with factual, up-to-date answers.
What exactly is DeepSeek V4?
It's the generation of DeepSeek models released on April 24, 2026, in two variants: V4-Pro (the most powerful) and V4-Flash (the fastest and cheapest). Both handle a million tokens of context, reason in two modes (with or without thinking), and are released as open source under the MIT license. It's the reference version today: when people talk about DeepSeek in 2026, they mean V4.
V4-Pro or V4-Flash: what's the difference?
V4-Pro activates 49 billion parameters per token and targets the hardest tasks: large coding projects, mathematical reasoning, analysis of massive documents. V4-Flash activates 13 billion, costs about 12 times less, and is more than enough for writing, translation, summarizing, and everyday code. Simple rule: start with V4-Flash, move up to V4-Pro only when you hit a complex task.
Which version of DeepSeek should I use?
For general chat use: V4-Flash, it's the default model and it's free. For demanding development or analysis of large documents: V4-Pro. The older versions (V3.1, V3.2, R1) hold no value for a new user unless you have a specific technical need. No need to overthink it.
Is DeepSeek free?
Yes, for the most part. The web version and mobile app are free, with no mandatory subscription. On the developer side, the API offers a starting credit, then bills per token at a rate well below equivalent closed models. Be wary of third-party sites that promise "free DeepSeek" and ask for a credit card: the real DeepSeek doesn't do that.
DeepSeek and data privacy: is it safe to use my data?
Important point. Regulators opened an inquiry as early as January 2025, and some countries suspended the app, though it has not been banned everywhere. With the official app, your data passes through servers in China, which raises real privacy questions under GDPR. The alternative: since DeepSeek is open source (MIT), your data stays with you if you use a self-hosted or Europe-hosted version. Don't put sensitive information into the consumer app.
Can you access DeepSeek from anywhere?
Yes, directly, without a VPN in most regions. One practical tip: avoid using a VPN or a disposable email address to create your account, or the confirmation code may never arrive. Access works normally from a standard browser.
The 30-second summary
What to remember about DeepSeek models
Confirmed on one side, rumor on the other. Don't mix them up.
Officially confirmed
- DeepSeek V4-Pro and V4-Flash launched on April 24, 2026, under the MIT license
- One-million-token context, two reasoning modes
- July 24, 2026 removed some old API identifiers, it is not a new model
- Open access, free web version and app
Unverified rumor
- No official date for a DeepSeek V5
- R2 has been denied by DeepSeek several times
- The V5 or R2 spec "leaks" circulating online are not confirmed
- Be wary of release dates announced without an official source
Official and verified sources
Check it for yourself
We'd rather give you the sources than ask you to take our word for it.
Explore further
Every page on DeepSeek models
Each version, each comparison, and each rumor, covered separately.
Get started
The best way to choose a model is to try it
Open the chat and test V4-Flash free on a real task: a snippet of code, a translation, a long document to summarize. You'll see in five minutes whether DeepSeek fits your needs, with no install and no credit card.