Will there be an open replication of DeepSeek v3 for <$10m?

MANIFOLD

Ṁ1kṀ4.7k

resolved Jul 10

Resolved

N/A

ALL

Background

DeepSeek-V3 is a large language model that DeepSeek claims was trained using 2.788 million H800 GPU hours, costing approximately $5.576 million in direct training costs. This is notably efficient compared to similar models like Llama 3.1 405B, which required over 11 times more GPU hours.

Resolution Criteria

This market will resolve YES if:

A team openly publishes reproducible or otherwise credible replication of DeepSeek-V3 with comparable performance
The compute cost is less than $10 million

The market will resolve NO if:

No successful replication is achieved by the resolution date
A replication is achieved but costs $10 million or more
A replication is achieved but there are no credible reports of the costs

Considerations

The cost of GPU compute will be calculated using December 2024 prices.

Technology

Open Source

Get

1,000

to start trading!

People are also trading

Will deepseek v5 release in september 2026?

4% chance

Will DeepSeek's next reasoning model be called R3?

1% chance

Will DeepSeek release a proprietary model for sale before EOY 2026?

15% chance

Will ANY DeepSeek model cause or materially enable a catastrophic risk by 2027?

3% chance

Did DeepSeek use a cluster of more than 5,000 H100s in 2024?

30% chance

Is there an American spy working at Deepseek?

44% chance

Will DeepSeek be banned in US before 2027?

5% chance

Will the first major multimodal DeepSeek release be video oriented and suited for robot control?

9% chance

Sort by:

@mods please re-resolve to N/A. The question did not specify that the replication has to be exact; the spirit of the question was about achievable compute efficiency, to figure out whether DeepSeek were truthful about the costs. Kimi k2 did not strictly match the architecture, but to the extent it’s close enough it is not clear whether to count it or not. Because of that, I don’t feel comfortable with a No resolution.

@ms you got it. thanks for following up!

@mods please resolve this market to "no", as an open-source reproduction of DeepSeek v3 did not happen in 2025.

There was a model (Kimi k2) which used the same architecture, but they changed it a touch for it did not strictly match DeepSeek v3

The reported cost for training Kimi k2 of 4.6 million is also unverified

https://www.yicaiglobal.com/news/kimi-k2-thinkings-reported-usd46-million-training-cost-isnt-official-moonshot-ceo-says

Moonshot AI’s CEO Says Reported USD4.6 Million Cost of Training Kimi K2 ‘Isn't Official’

@Eternal @mods please re-resolve to N/A. The question did not specify that the replication has to be exact; the spirit of the question was about achievable compute efficiency, to figure out whether DeepSeek were truthful about the costs. Kimi k2 did not strictly match the architecture, but to the extent it’s close enough it is not clear whether to count it or not. Because of that, I don’t feel comfortable with a No resolution.