Kimi K3 (thinking) 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 (thinking) leads overall, 89.2 to 84.2.
The verdict
Kimi K3 thinking is #12 and the new DeepSeek V4 Flash 0731 is #34, 89.25 against 84.23. Six and a quarter points. The gap is widest on length adherence, 91.10 against 83.57, and YouTube structure, 87.83 against 81.06. DeepSeek's hooks are the bright spot at 87.55, within three of Kimi, so openings land. What it does not do is sustain: continuity 82.02 against 88.28. Both are open weights, so this is a straight quality-versus-price call rather than a licensing one, and the price gap is enormous: half a cent a script against 19 cents. Forty three times cheaper. For six points on a draft you were going to edit anyway, that is a genuinely hard argument to beat at volume. For a script you want close to final, Kimi is worth the money.
Pick Kimi K3 (thinking) when the draft needs to hold its length and flow with minimal editing.
Pick DeepSeek V4 Flash 0731 for volume. Both are open, but this one is forty three times cheaper with comparable hooks.
Metric by metric
Blue bars: Kimi K3 (thinking). Orange bars: DeepSeek V4 Flash 0731. Same 0–100 scale; the bold bar wins that metric.
Everything else that differs
| Kimi K3 (thinking) | DeepSeek V4 Flash 0731 | |
|---|---|---|
| Overall / 100 | 89.2 | 84.2 |
| Writing Elo | 2416 | 1955 |
| Run-to-run spread (± overall std) | 1.650 | 9.240 |
| Cost per script (USD) | 0.187 | 0.005 |
| Avg latency (s) | 267.2 | 289.0 |
| Open weights | Yes | Yes |
Full scorecards: Kimi K3 (thinking) · DeepSeek V4 Flash 0731. How scoring works: methodology.
← All comparisons