Open-weight & challenger labs
DeepSeek
A Chinese lab that moved past its well-known V3/R1 naming to a new V4 generation in 2026, generally released with open weights.
Checked September 19, 2026
No lab publishes a real roadmap
So this page keeps three things apart on purpose. What has already shipped is dated. What is coming is listed only where the lab itself has said it in public, with the source named. Everything else sits under “What to watch” and is my own read of their shipping pattern — informed guessing, not news. If a date matters to you, check the lab's own page linked above.
What they are
A Chinese lab that made the biggest single dent in the assumption that frontier models must be expensive to train and closed to inspect.
It has moved past the V3/R1 naming most people know, and it generally publishes weights for its flagship — which very few labs do.
What they currently ship
Every model in the directory entry, with its free-tier status.
- DeepSeek-V4-Pro— Current flagship for general and agentic tasks with adjustable reasoning effort — open weights on Hugging Face, also available paid via API.Free, open-weight
- DeepSeek-V4.1-Flash— Smaller, faster sibling with native image understanding.Free, open-weight
What you get for free
The honest version, in one paragraph.
Free in the way that matters most: the flagship weights are published, so you can run it yourself or use a third-party host. There is no free hosted tier from DeepSeek itself — API use is paid — but nothing is locked away from you.
Track record
Things that actually happened, oldest first, each with a date.
- December 2024DeepSeek-V3 ships with open weights and unusually low reported training cost.
- January 2025DeepSeek-R1 brings open-weight reasoning models into the mainstream conversation.
- 2026The V4 generation arrives: V4-Pro as flagship, V4.1-Flash as the smaller, faster sibling with native image understanding.
Confirmed and announced
Only what the lab has stated in public. Nothing here rests on a rumor.
DeepSeek-V4-Pro is published with open weights on Hugging Face as well as being available paid via API.
Announced by the lab
What to watch
My own read of where this lab is heading, based on how it has shipped so far. Not announcements — treat these as questions to check, not facts.
- — Whether the open-weight-flagship policy holds for the next generation, or narrows as the models get more valuable.
- — Availability and export restrictions, which affect who can use these models more than price does.
- — How quickly third-party hosts pick up each release, since that is what makes them usable without your own GPUs.
Their ethics and safety policies
What DeepSeek has published about safety and ethics, the documents themselves, what those documents leave out — and my own read, kept separate.
Looking for which tool to actually use? See AI Models Compared