How good is Kimi K3 at writing fiction?

Where it stands against the board

The headline is generational proof: K2.6's two board-documented ailments — the field's highest simile density and whole-passage self-copying mid-chain (97.6% cross-round 12-gram repetition at r9 of the archived chain) — both vanished in K3's same-protocol retest: repetition peaked at just 6.8% (xuanhuan r10, ≈0% elsewhere), and stock-phrase density converged to 2.02‰ on xuanhuan and 0.63‰ on romance (below the 1.0‰ source baseline). Judges credited it with “perfectly cloning the source's short paragraphs, standalone onomatopoeia and '~~~' punctuation habit” — micro-texture only gpt-5.6-terra had replicated in the nine-model round.

Against the two genre champions (the DeepSeek pair) there is only a same-ruler comparison, not a shared blind duel: on xuanhuan K3's composite distance of 0.388 lands between V4 Flash (0.356, that genre's blind champion) and V4 Pro (0.454); on romance its 0.396 doesn't reach V4 Pro's 0.280. Who beats whom has no exam paper yet — that duel is on the schedule.

Its standing as the current continuation strongman has corroboration: Qwen3.8-Max-Preview, released the same month, challenged it under the identical protocol and lost 3:20 (1:10 xuanhuan / 2:10 romance). Later incremental duels use K3 as the reference strongman.

The warts have ballots too: ~85% of dialogue in half-width quotes (the same stubborn habit as glm-5.2, a constant immersion break in Chinese typography); first-person density at 16.2‰ — 1.8× the source — on first-person period prose; thinking permanently on, eating 81-86% of completion tokens at full price, with 37-54s waits per round.

Test-condition disclosure (hosted models are moving targets)

Model under test: kimi-k3 (released 2026-07-16)

Evaluated: 2026-07-18 · Access channel: Moonshot official API (thinking always on)

Review format: paired double-blind verdicts (two flipped mappings per book against position bias), not the nine-model full ranking

Protocol matches the main board item by item (same two books · same 55% anchor · same D2 directive · 20-round chains · temperature 0.7), max_tokens=2800; paired baselines = the archived K2.6 romance chain plus a freshly run K2.6-D2 xuanhuan chain to eliminate the directive variable.

Honest limits

Kimi K3 has not entered the nine-model same-protocol full-ranking review, so the board's rank column does not apply to it — this page publishes only ballot-backed paired duels and invents no rank. When it joins the full ranking depends on the next full-board rerun.

Paired blind duels ran only against K2.6 (same-vendor predecessor) and Qwen3.8 — never in the same blind room as the DeepSeek pair, so its record can't be hard-compared against the nine-model board ranks (different review formats: full ranking there, paired verdicts here).

Single book, single anchor, single chain per book (n=1), matching the main board's per-model conditions; the two invalid xuanhuan ballots were a claude judge refusing JSON output and writing its own continuation instead — excluded as recorded.

What 10,000 characters cost

$0.81 list price: in $3/M · out $15/M (models.dev snapshot 2026-07-24)

Using the app's continuation recipe: one segment ≈ 400 chars = 8k input + 550 output tokens; 10k chars ≈ 25 segments; no cache discount. For between-model comparison only.

Keep going