DeepSeek's training-cost figure excluded its research spend

Two models, two reports, three weeks apart

DeepSeek released its DeepSeek-V3 Technical Report on 27 December 2024, describing a 671-billion-parameter mixture-of-experts model with 37 billion parameters active per token, and its DeepSeek-R1 paper on 22 January 2025, describing a reasoning model trained largely through reinforcement learning. Both were released as open weights alongside their papers, distinguishing the announcements from a closed-model launch post: the underlying claims can, in principle, be checked against the released model rather than taken only on the paper's word.

What the training-cost figure covers

The V3 report states a specific, itemised cost: 2.664 million H800 GPU hours for pre-training, plus 119,000 for context-length extension and 5,000 for post-training, totalling 2.788 million GPU hours, which the paper converts to about $5.576 million assuming a $2-per-GPU-hour rental rate. The report is explicit about what this figure excludes: it states the costs 'include only the official training of DeepSeek-V3, excluding the costs associated with prior research and ablation experiments on architectures, algorithms, or data.' That boundary comes from the paper itself, not an outside estimate, and it means the widely repeated headline figure describes one final training run rather than the cumulative cost of the research programme that produced it.

How R1 was trained, and what the comparison to o1 does not establish

The R1 paper distinguishes two models: DeepSeek-R1-Zero, trained purely by reinforcement learning on the DeepSeek-V3 base with no supervised fine-tuning step, and DeepSeek-R1, which adds a 'cold start' of curated long chain-of-thought examples before reinforcement learning to fix R1-Zero's poor readability and language mixing. The paper reports R1 performing on par with OpenAI's o1-1217 checkpoint on maths tasks, while o1-1217 leads on one engineering-coding benchmark and the two are comparable on another. The paper also states plainly that 'accessing the OpenAI-o1-1217 API is challenging in mainland China,' so DeepSeek reports o1-1217's scores from OpenAI's own published results rather than from a benchmark run DeepSeek conducted itself — a comparison built from two different labs' testing conditions, not one controlled evaluation. The project's GitHub repository licenses the code and released weights under MIT terms, while stating that each smaller distilled model carries the licence of the base checkpoint it was distilled onto — Apache 2.0 for the Qwen-based variants, Meta's Llama licences for the Llama-based ones — rather than a single blanket licence for the whole R1 family.

• Does the training-cost figure being cited include or exclude the research and failed-experiment costs the paper itself says are excluded?
• Which of R1's reported benchmark comparisons come from a run DeepSeek conducted directly, and which are taken from another lab's self-reported numbers?
• Do the distilled smaller models carry the licence of the base checkpoint they were distilled onto, or DeepSeek's own terms?

Open weights make a claim easier to test than a closed launch post, but easier is not automatic: a training-cost figure and a cross-lab benchmark comparison still carry the boundaries their authors set for them, and both papers here state those boundaries explicitly enough to be read rather than assumed.