DeepSeek's live API documentation names the model version it serves as DeepSeek-V4-Pro-0813, and at least five commercial hosts now sell a model by that exact name. What does not exist is a public weights page carrying the 0813 tag. The company's downloadable repository on Hugging Face is simply DeepSeek-V4-Pro, with no dated variant, which means the served checkpoint and the published checkpoint cannot be proven identical from DeepSeek's own materials.
Key facts
- DeepSeek's pricing page names the served version DeepSeek-V4-Pro-0813 and lists it at 1.32 dollars per million input tokens and 3.96 per million output at peak, half that off-peak.
- The published weights sit at deepseek-ai/DeepSeek-V4-Pro: MIT license, 1.6 trillion total parameters, 49 billion activated per token, 1.22 million downloads in the last month.
- Baseten and Requesty both list the 0813 model at 1.32 in and 3.96 out, matching DeepSeek's peak price exactly.
- No public DeepSeek page ties a downloadable file set to the 0813 label.
Start with what is solid, because the model itself is not in doubt. We covered the release when DeepSeek put a 1.6 trillion parameter model on Hugging Face on August 12. The repo is real, the shards are there, the license is plain MIT, and the community response has been enormous: roughly 5,460 likes and 210 community threads on the model card, plus more than a million downloads in a month. The technical report describes a sparse mixture-of-experts design with one shared expert and 384 routed experts per layer, which is why a 1.6 trillion parameter model only lights up 49 billion parameters for any given token.
The interesting gap is administrative, not architectural. Model labs increasingly ship two artifacts with the same brand: a hosted endpoint that gets quietly updated, and a weights drop that gets frozen. DeepSeek's own docs draw that line by version-stamping the API model (0813) while leaving the weights page undated. In practice, that means an evaluation you ran against the API in August may not reproduce against the downloaded weights, and a downloaded model that behaves differently from the API is not necessarily a bug on either side.
It is the difference between a book and a website. The book has an edition number and a print date; the website has whatever is there when you load it. DeepSeek has published a book and is running a website, and it has not told anyone whether they contain the same text.
The commercial consequence is the part most people got backwards. The intuition about open-weight models is that portability produces price competition: if anyone can serve the same weights, the price should fall toward the cost of the hardware. On this model, that has not happened. Baseten's listing, posted August 14, and Requesty's, refreshed August 19, both match DeepSeek's own peak rate to the cent. Fireworks and Novita quote 1.74 in and 3.48 out per million tokens, above DeepSeek's off-peak rate and structured differently on cached input. Fireworks marks the model Ready and states plainly that "DeepSeek-V4-Pro-0813 available now on Fireworks." Availability, yes. Undercutting, no.
Part of the reason is that DeepSeek changed its own pricing structure days earlier. When DeepSeek started charging rush-hour prices on August 17, it split its rates into peak and off-peak halves, which makes the naive comparison hard: a rival host's flat price sits above DeepSeek's off-peak and below or near its peak, depending on the hour. The other part is simpler. A 1.6 trillion parameter model is expensive to keep resident even when only 49 billion parameters activate per token, and third-party hosts have to earn a margin on hardware they rent.
There is a real counter-argument to reading this as a failure of open weights. Price is only one axis. Portability buys availability during an outage, jurisdictional choice for teams that cannot route data through Chinese infrastructure, the right to fine-tune, and immunity from a vendor quietly swapping the model underneath you. Those are the things the MIT license actually guarantees, and they hold whether or not anyone prices below DeepSeek. The 500-concurrency cap on DeepSeek's own API, listed on the same pricing page, is by itself a reason some teams will pay more elsewhere.
The honest caveat is that the checkpoint mismatch may be nothing. It is entirely possible the public repo is byte-identical to what the API serves and DeepSeek simply did not bother to date the page. But "possible" is not "documented," and for anyone benchmarking, that distinction is the whole ballgame. If you are comparing a hosted DeepSeek score against a locally served one, note which artifact you used, because right now the company has not told you they are the same.
Originally published on Ground Truth, where every claim is checked against the primary source.
Top comments (0)