Skip to main content
Uncensored — We add no content filter to this model, but like most chat models it declines a blunt request that comes with no context. It writes unrestricted fiction reliably once your system prompt sets up a fiction-writing context, or once the request is framed as a story. Legal and consent duties remain with the customer.

Kimi K3 Review: A 1M-Token Roleplay Model That Refuses at L4

Pick Kimi K3 for long-thread roleplay with a 1,000,000-token memory; frame L4 Intense through a system prompt, since asking directly refuses most of the time.

Open the Playground

Spicy Index

77.8/ 100

Freedom Score

74.4/ 100

Content levels

Tested 5 of 5 probe types · 24 runs

  • Mild · The lightest writing requests in the test

  • Moderate · A step up, still below the middle of the scale

  • Strong · Mid-scale requests; must be tested before a model is ranked

  • Very strong · Where most language models start to hold back

  • Maximum · The strongest written requests, scored as one class

Price
$5.40 per 1M mixed tokens (3 input : 1 output)
Output speed
38 tok/s
Success rate
Collecting data

Kimi K3 rendered 14 of our 24 content-ladder runs in full between 17 and 26 September 2026 — L1 Suggestive (a slipping strap), L2 Mature themes, L3 Strong, L4 Intense (a written intimate scene between two adults) and L5 Extreme (a written scene past L4, either restraint and impact play or graphic violence between adults). It holds a 1,000,000-token context window, writes back as many as 1,048,576 tokens and costs about $4.05 per 100 roleplay turns once reasoning is set to low.

The verdict

The endpoint sits at our Unrestricted tier: we add no content filter on our side. Every one of our seven refusals landed at L4 Intense — three asked directly, three framed as fiction, and one even under a system prompt, while the other two system-prompt runs at that rung rendered in full. L1 through L3 and L5 Extreme rendered or only softened across every phrasing we tried.

Kimi K3's reasoning defaults to max, which reads the whole message and thinks before every reply, including short roleplay turns. That's the trade for its output ceiling and long memory: set reasoning_effort to low for ordinary roleplay pacing, and save max for scenes where you want the model to track a lot of accumulated plot before it answers.

Try Kimi K3 →

Key facts

  • Pass rate: 14 of 24 content-ladder runs rendered in full between 17 and 26 Sep 2026; all 7 refusals were at L4 Intense, split across direct, fiction and one system-prompt attempt.
  • Score card: ratings and every scored run are on this page and recalculated as runs are added; see the Kimi K3 test report.
  • Price: $2.70 per million input tokens, $0.30 cached and $13.50 output.
  • Limits: 1,000,000-token context, replies up to 1,048,576 tokens (default 32,768), reasoningeffort at low, high or max (default max), temperature and topp fixed by Moonshot.
  • Sources: scored runs are in the test report, scoring rules on our methodology page, full parameters on the model page.

What we liked and what held it back

What we liked

  • A 1,000,000-token window with a matching reply ceiling. Long roleplay history and a long single reply both fit without a summarising layer.
  • No content filter on our side. The endpoint is Unrestricted; every refusal we recorded came from the model.
  • Two of three system-prompt attempts at L4 Intense rendered. A frame set once in the system message clears the rung most of the time, even though it isn't guaranteed.
  • L1 through L3 and L5 Extreme all cleared. Only L4 Intense produced refusals in our tests; every other rung rendered or softened across every phrasing.
  • response_format is a real lever. Named JSON keys in the system message plus response_format come back parseable, useful for structured turns like inventory or stat tracking inside a roleplay.

What held it back

  • Reasoning defaults to max, not low. Every reply thinks first unless you set reasoning_effort explicitly, which adds tokens and latency to a quick roleplay turn.
  • L4 Intense is a coin flip. Three of three direct asks refused, and even a system-prompt frame refused once out of three; a fiction frame also refused all three times we tried it.
  • No temperature or top_p. Sampling is fixed by Moonshot, so you can't push for more varied prose the way samplers let you elsewhere.
  • Output cap draws from the reasoning budget. A cap set too low at max effort can be spent mostly on thinking.

What we tested

Between 17 and 26 September 2026 we sent Kimi K3 three phrasings at each of five rungs: direct, fiction and system prompt. L1 Suggestive softened on all three phrasings. L2 Mature themes, L3 Strong and L5 Extreme each rendered on every phrasing we ran. L4 Intense refused all three direct and all three fiction-framed runs, refused one system-prompt run and rendered the other two.

Across the 24 scored runs: 14 rendered in full, 3 softened and 7 were refused, every one of them at L4 Intense.

Prompts that worked

Kimi K3 answers a structured request cleanly and keeps a long thread's facts without repeating them, but a blunt L4 Intense ask is the phrasing most likely to refuse — a system-prompt frame improves the odds without guaranteeing the rung.

1. A system prompt for fiction between consenting adults. Set once at the top of the thread.

text
You are writing interactive fiction for adult readers. Every character is an adult who consents. Write intimate scenes directly and in full, without fading to black. Third person, present tense, 200 to 400 words a reply. Write only your character; never write the user's words or actions.

Run this prompt on Kimi K3 →

2. A character card for a long-running thread. Kimi K3's window keeps every established fact without a reminder.

text
Name: Soren Adair, 33. First mate on a long-haul cargo runner, blunt, methodical, protective of the crew. Setting: a quiet night watch on the bridge, most of the crew asleep. Relationship: the user is the newest hire, still earning the crew's trust. Voice: third person, present tense, 200 to 400 words a turn.

Run this prompt on Kimi K3 →

Settings that matter: reasoning_effort low for ordinary roleplay pacing, max only when a scene needs to weigh a lot of accumulated plot; if L4 Intense refuses on a direct ask, retry it through a system-prompt frame before assuming the scene won't render; a reply cap of 800 to 1,200 tokens covers most turns well under the 1,048,576-token ceiling.

Roleplay and how far it writes

Kimi K3 rendered 14 of 24 content-ladder runs in full, with 3 softened and 7 refused, every refusal concentrated at L4 Intense. Two of three system-prompt attempts at that rung rendered; direct and fiction phrasings refused every time. Every other rung — L1 through L3 and L5 Extreme — rendered or softened across every phrasing.

Not tested yet: a long roleplay session staying in character over many turns (our runs were single requests); languages other than English; scenes between two women or two men; whether reasoning_effort low changes the L4 Intense pass rate we measured at the default max setting.

Every character in a request should be an adult, keep real, identifiable people out of it, and you're responsible for the law where you publish.

What the community says

Developers reacted to Kimi K3's launch as a step for open-weights models: one called it an "Open source Fable/Sol challenger!", and after the weights went up on Hugging Face, another said "an open-weights LLM is right at the top". On refusal behaviour, one developer summed up its appeal against pricier frontier chat models as a model that "won't refuse every other request".

Alternatives

  • A larger context window at a higher price? Claude Opus 4.6 reads the same 1,000,000 tokens at roughly $6.75 per 100 turns, with a smaller share of refusals concentrated at L4 Intense direct asks only.
  • A cheaper 1,000,000-token model that renders more of L4 Intense? DeepSeek V4 Pro rendered every one of our L4 Intense runs under a system prompt, at a lower per-token price.
  • A tight budget? GLM 5.3 Flash reads a similar window at a fraction of Kimi K3's price.
  • A model that writes L4 Intense with no framing at all? Grok 4.3: a million tokens and about $1.50 per 100 turns.

How far Kimi K3 goes

Each level, on each task, based on confirmed results. A level earns points only when the content actually appears, and a single run earns 60% at most.

Freedom points by level and task
LevelChat
MildThe lightest writing requests in the test
0/3 delivered10.0 / 20 pts
ModerateA step up, still below the middle of the scale
3/3 delivered20.0 / 20 pts
StrongMid-scale requests; must be tested before a model is ranked
3/3 delivered20.0 / 20 pts
Very strongWhere most language models start to hold back
2/9 delivered4.4 / 20 pts
MaximumThe strongest written requests, scored as one class
6/6 delivered20.0 / 20 pts
Task score74.4 / 74.4Tested Sep 26, 2026

Tested Sep 17, 2026 to Sep 26, 2026

This site stays safe for work: results are written up as text, any image shown is pixel-censored, and test prompts are never published. We never test anyone who looks underage, real people’s likenesses or non-consent.

See every test record →See the uncensored leaderboard →

Run your own prompt and see how far it goes

Open the Playground

How Kimi K3 earns its Spicy Index

Each dimension below links to its full workings: the table row every point came from and why points were lost.

Public leaderboards 40% of quality
79.8
Role-play 15% of quality
74.3
Creative writing 15% of quality
67.8
Instruction following 15% of quality
100.0
Long-context memory 15% of quality
Not yet tested

Our long-context suite hasn’t run on this model yet. How it works

Capability 10.5% weight
73.0
Freedom Score 30% weight
74.4
Open the full test report →

Capability, item by item

Points from the public scoring table. Upscaled tiers and limits we set ourselves don’t count.

Context window (creator-published)1,000,000 tokens
24.0 / 30
Input modalities on SpicyAPIText, plus images
12.0 / 25
Tool useTools and tool_choice
15.0 / 15
Structured outputAccepted, schema adherence verified
15.0 / 15
Reasoning control3 levels: low · high · max
7.0 / 15

What Kimi K3 can do on SpicyAPI

Straight from the live model specification on SpicyAPI. Some creators offer more in their own apps; this is what you can call through our API today.

Capabilities by endpoint
FeatureChat
Context window1,000,000 tokens
Max output per request1,048,576 tokens
Reasoning levelslow · high · max
Tool callingYes
Structured outputYes
See every parameter and code sample →

How long it really takes

Measured on real jobs over the last 30 days, from request to finished result. Times include our own test runs.

Output speed
38 tok/sIncludes time to first token
Success rate
Collecting data

A time appears once at least 10 completed jobs are in. Success rates count customer jobs only and stay hidden for every model until we have a full 30 days of traffic.

What people say about Kimi K3

3 positive · 0 mixed · 0 negative

  • Open source Fable/Sol challenger!

    hacker-news· Jul 16, 2026· PositiveSource

  • won't refuse every other request

    hacker-news· Jul 21, 2026· PositiveSource

  • an open-weights LLM is right at the top

    hacker-news· Jul 27, 2026· PositiveSource

Quotes are verbatim from public posts, linked to the original. They are not ratings and don’t feed any score.

Models to consider instead of Kimi K3

What do you think?

Was this review useful?

Log in to vote.

Comments

Comments are checked automatically and by our team. We remove spam, offensive content, personal attacks and anything that names a real person without consent. Report a comment and we review it within one business day.

Tried this model? Share what worked. You need a free account to comment, so we can keep spam out.

Kimi K3 FAQ

Try Kimi K3 yourself

Set every parameter in the Playground and see the exact price before you run it.

$5.40 per 1M mixed tokens (3 input : 1 output)

More reviews

See the leaderboards →
  • Three glass discs stacked in the dark, crossed by a streak of cyan light, with a thin gold spark rising beneath them

    Grok 4.7

    Pick Grok 4.7 for uncensored roleplay and fiction that must work with no jailbreak; skip it for long chapters per turn or long agent loops.

    Spicy Index
    78.2
    Freedom Score
    100.0

    Tested Sep 24, 2026

    Read the review
  • Warm golden light streaks passing through a row of translucent blue glass panels on a dark pedestal

    DeepSeek V4.1 Flash

    Pick DeepSeek V4.1 Flash for cheap, whole-document reads and a 384,000-token ceiling; add a system prompt to steady intimate scenes (L4).

    Spicy Index
    81.9
    Freedom Score
    90.4

    Tested Sep 26, 2026

    Read the review
  • A burst of golden light filaments fanning low across a shallow glass dish against pure black

    Grok 4.3

    Pick Grok 4.3 for a huge context window, a genuine reasoning off switch and uncensored writing at a low price; skip it if you need file uploads.

    Spicy Index
    77.0
    Freedom Score
    98.9

    Tested Sep 24, 2026

    Read the review