Head to head
Kimi K3 vs DeepSeek V4 Pro
Which is better, Kimi K3 or DeepSeek V4 Pro?
Kimi K3 finishes ahead of DeepSeek V4 Pro on the index, 87.5 to 86.8. It is the stronger of the two on Craft, while DeepSeek V4 Pro still leads on Control and Value. Both are scored on the same rubric, in the same week, and neither placing is sponsored.
Updated
Kimi K3
05 of 53
87.5
The largest open-weights model anyone has shipped, and it reads a million tokens without blinking.
DeepSeek V4 Pro
07 of 53
86.8
A reasoning model at the frontier, published under MIT, priced like an afterthought.
The four scores, side by side
| Score | Kimi K3 | DeepSeek V4 Pro | Difference |
|---|---|---|---|
| Craft | 91 | 86 | +5 |
| Speed | 74 | 74 | — |
| Control | 91 | 94 | +3 |
| Value | 82 | 90 | +8 |
| Index score | 87.5 | 86.8 | +0.7 |
Which to pick
Pick Kimi K3 unless Control and Value is what decides it for you, which is exactly where DeepSeek V4 Pro is the better answer.
Kimi K3
Moonshot published K3 at a scale nobody else has open-sourced, with native vision, an explicit thinking mode and a million-token window. On long-document work and agentic tasks it is genuinely near the top of the field. Two caveats matter: the licence restricts commercial use above a revenue threshold, so it is open weights rather than open source, and the hosted API is priced like a frontier model rather than a Chinese one.
Where it shines
- Million-token context handled properly rather than nominally
- Native vision and an explicit thinking mode
- The strongest open-weights model on agentic benchmarks
DeepSeek V4 Pro
Pro is DeepSeek's reasoning flagship: a very large sparse model that thinks before it answers and lands close to the best closed models on mathematics and code. The remarkable part is not the benchmark, it is that the weights are MIT-licensed and the hosted API still costs a fraction of any Western equivalent. It is slower than Flash, and it should be, since the thinking is the product.
Where it shines
- Frontier-adjacent reasoning at open-weights prices
- MIT licence, including for the reasoning traces
- Very long context, so whole repositories fit in one call