K Kimi K3 (thinking)
Moonshot (Kimi) · open weights · writing benchmark
The short version
Kimi K3 (thinking) sits at #12 of 130 on ToneBench, in the upper tier, with a writing Elo of 2416 and an overall score of 89.2 out of 100. We measured it by having it write all 9 of our real YouTube scripts, five times each, then scoring every draft blind against our own finished versions. Here is how it shook out.
Within Moonshot (Kimi) it ranks 2nd of 6. Kimi K3 is the family's top writer here, about 21 Elo ahead, so this config trades some quality for whatever it saves you in cost or speed.
Among open-weights models, it comes in 2nd of 47. That is a genuinely strong showing for an open-weights model on a voice-and-tone task, which is usually where the closed frontier still pulls ahead.
One thing that stands out: it is unusually consistent. Across its 45 drafts the overall score only moves ± 1.6 points run to run, versus a board median of ± 3.8. What you get on the first try is close to what you get every try.
What it does best is Hook: it ranks 6th on that dimension at 90.2, well above the board average. Its softer spot is Anti-Slop (91.6, 31st), which is the thing to watch if that metric matters most for your use.
It held steady across all 9 scripts, with its best run on the founder announcement / personal origin story (90.1). Consistency like that is underrated, honestly, since it means fewer surprises when you actually put it to work.
On cost, it runs about $0.187/task. Claude Opus 5 (low) scores higher for less money, at $0.115/task, so on pure value this config is not the frontier. Still, it beats 5 models that cost noticeably more.
Skill profile
The shape is the story: the further a corner reaches, the stronger that dimension. Biggest edge over the board: Length Discipline (+23.9); smallest edge: Anti-Slop (+8.8).
Per-metric scores
Nine writing dimensions, each scored 0–100 and blended by the editorial weight shown. Rank is against all 130 current-ranked models. The thick bar is this model; the thin lines above and below are the current field's best model (darker) and average (lighter) on the same scale. The small ± number is its run-to-run variation.
Per-article scores
The same 9 real scripts every current-ranked model writes, scored individually. Different formats stress different skills.
Consistency: run-to-run variability
Every cell on this page is the mean of 5 independent runs per script; the ± numbers are the run-to-run standard deviation. Overall, Kimi K3 (thinking) varies by ± 1.6 points between runs versus a board median of ± 3.8, so it is noticeably steadier than the typical model here.
Its steadiest is YouTube Structure (± 1.9 vs ± 3.6 board median), which you can rely on run after run.
| Metric | This model | Board median | Verdict |
|---|---|---|---|
| Tone & Voice ? | 89.3 ± 1.5 | ± 2.2 | typical |
| Writing Craft ? | 89.3 ± 1.4 | ± 2.0 | typical |
| Substance & Value ? | 89.2 ± 1.7 | ± 2.6 | typical |
| Flow & Emotion ? | 88.3 ± 1.7 | ± 2.6 | typical |
| YouTube Structure ? | 87.8 ± 1.9 | ± 3.6 | steadier than most |
| Hook ? | 90.2 ± 1.7 | ± 3.0 | steadier than most |
| Length Discipline ? | 91.1 ± 6.2 | ± 8.6 | typical |
| Anti-Slop ? | 91.6 ± 2.1 | ± 2.6 | typical |
| Visual Cues ? | 87.7 ± 2.5 | ± 4.7 | steadier than most |
Measured, not modeled
Numbers we log on every run and rarely talk about. None of these affect the writing scores; cost and latency are informational.
What each judge scored it
The published overall of 89.2 is the consensus of three family-disjoint judges scoring the same 45 stored drafts blind with the identical rubric. The highest and lowest judge differ by 5.5 points on its overall. The three judges essentially agree on this model. How the panel works: methodology.
How we ran Kimi K3 (thinking)
- Accessed viaOpenRouter · model page ↗
- Exact model / route id
moonshotai/kimi-k3atreasoning_effort: high - Generations5 independent runs × 9 scripts = 45 drafts, stored immutably after generation
- Run dates2026-07-29 → 2026-08-11
- Judges3-judge panel: Claude Opus 5, high effort (Anthropic) + DeepSeek V4 Flash (OpenRouter) + GPT-5.6 Sol, medium reasoning (Codex CLI): every draft scored independently by all three with the identical rubric, blind to model identity, evidence quotes required before each metric score; published score = per-metric consensus mean, verified with a Bradley-Terry refit (methodology)
- Samplingprovider-default sampling, no fine-tuning, no system-prompt changes; every current-ranked model gets the identical brief, style contract and research packet
- What stays privatethe exact rubric and any non-public reference material; source-publication timing is disclosed for exceptions (why)
Frequently asked questions
How good is Kimi K3 (thinking) at writing?
On ToneBench it ranks #12 of 130 with a writing Elo of 2416 and an overall score of 89.2/100. That score comes from writing our 9 real YouTube scripts five times each and scoring every draft blind against our own finished versions across nine writing dimensions.
Is Kimi K3 (thinking) the best Moonshot (Kimi) model for writing?
Not quite. Within Moonshot (Kimi) it ranks 2nd of 6; Kimi K3 is the family's best writer here.
Is Kimi K3 (thinking) good value for the money?
It costs about $0.187/task. Claude Opus 5 (low) scores higher for less, so it is not the value pick.
What are Kimi K3 (thinking)'s strengths and weaknesses?
Its strongest dimension is Hook (6th on the board, 90.2). Its weakest is Anti-Slop (31st on the board, 91.6). The full nine-metric breakdown is on this page.
Does Kimi K3 (thinking) write some formats better than others?
It was fairly even across all 9 scripts, topping out on the founder announcement / personal origin story (90.1).
How consistent is Kimi K3 (thinking) between runs?
We run every script 5 times. Kimi K3 (thinking)'s overall score varies by about ±1.6 points between runs, versus a board median of ±3.8. That makes it one of the steadier models we test. The full per-metric variability table is on this page.
How was Kimi K3 (thinking) evaluated?
Via OpenRouter using the exact model/route id moonshotai/kimi-k3 at reasoning_effort high, run on 2026-07-29. 5 runs per script, provider-default sampling, no fine-tuning; every draft scored blind with a fixed rubric by a three-family judge panel. Full details in the 'How we ran it' section and the methodology.
Is Kimi K3 (thinking) open source?
Yes, it is an open-weights model. Among open-weights models it ranks 2nd of 47 for writing.