Kimi K3 vs DeepSeek V4 Flash 0731
Head-to-head on the Towards AI writing benchmark: same 9 real YouTube scripts, five runs each, scored blind by a three-family judge panel. Kimi K3 leads overall, 89.4 to 84.2.
The verdict
The best open-weights model on the board against the cheapest one worth using, and the gap is honest: Kimi K3 at #8 with 89.39 overall, DeepSeek V4 Flash 0731 at #34 with 84.23, five points, intervals nowhere near each other. Kimi is simply the better writer everywhere I look: tone 89.61 against 85.44, continuity 88.28 against 82.02, YouTube structure 88.39 against 81.06. DeepSeek's hooks at 87.55 are the one place it plays in Kimi's league. So why is this even a page? Price. DeepSeek runs about half a cent a script against Kimi's 26 cents, roughly fifty times cheaper, and the 0731 build just climbed 49 places in one release. Five points is a real quality gap; fifty times is a real budget gap. If the script ships to my channel, I pay for Kimi. If I'm bulk-drafting and a human owns the rewrite anyway, DeepSeek's math is unbeatable.
Pick Kimi K3 when the draft has to stand on its own: it wins every writing metric by a wide margin and is still open weights.
Pick DeepSeek V4 Flash 0731 for volume drafting with a human editor: about fifty times cheaper, respectable hooks, and improving fast.
Metric by metric
Blue bars: Kimi K3. Orange bars: DeepSeek V4 Flash 0731. Same 0–100 scale; the bold bar wins that metric.
Everything else that differs
| Kimi K3 | DeepSeek V4 Flash 0731 | |
|---|---|---|
| Overall / 100 | 89.4 | 84.2 |
| Writing Elo | 2438 | 1955 |
| Run-to-run spread (± overall std) | 1.570 | 9.240 |
| Cost per script (USD) | 0.263 | 0.005 |
| Avg latency (s) | 367.1 | 289.0 |
| Open weights | Yes | Yes |
Full scorecards: Kimi K3 · DeepSeek V4 Flash 0731. How scoring works: methodology.
← All comparisons