Se connecter
Kimi K3 logo

Kimi K3

Moonshot AI · Sorti juil. 2026

Voir les tarifsComparer les modèles
Conditionnel

Moonshot AI's open-weight MoE model (Kimi K3 License, weights on Hugging Face) offers a 1M-token context at $3/$15 per 1M tokens and holds #1 on Arena WebDev (preliminary), but still trails Fable 5 and GPT-5.6 Sol on precise execution, and serving the weights needs 64+ accelerators. The rating stays conditional on legal risk: on July 22, 2026 the White House OSTP accused Moonshot of distilling Fable 5 and Treasury threatened sanctions. We have not re-verified that dispute's status this run.

Est-ce fait pour vous ?

Recommandé pour

  • Long-horizon autonomous coding — MiniTriton compiler, 48-hour chip design demonstrations
  • Front-end and 3D work with image feedback — Arena WebDev #1 (1679, preliminary)
  • Multi-agent orchestration — K3 Swarm Max variant for large-scale parallel processing
  • Large-context knowledge work — 1M window for whole-repo and multi-document research

Déconseillé pour

  • Precise-execution tasks — still trails Fable 5 and GPT 5.6 Sol
  • Self-hosting on modest hardware: the weights are downloadable from Hugging Face under the Kimi K3 License, but serving them needs 64+ accelerators
  • Cost-sensitive workloads — AA measured high output verbosity driving up task costs

Tarifs

Entrée

$3.00 / 1M

Sortie

$15.00 / 1M

Contexte

1M ctx, $0.30 cached. Weights live (Jul 27).

Voir tous les tarifs

Tests de performance

TestScoreSource
DeepSWE67.5 Source
Terminal-Bench 2.188.3 Source
FrontierSWE81.2 Source
Program Bench77.8 Source
SWE Marathon42.0 Source
Arena WebDev1679 (#1, preliminary) Source
AA Intelligence Index57 Source

Historique des verdicts

30 sept. 2026
conditional -> conditional (summary expanded from 28 words, stale investigation framing removed): Rating unchanged. The standing summary was 28 words and described a live White House investigation with no pricing, weights or benchmark context. Rewritten from stored, sourced children; the dispute is attributed to the July 22 record and flagged as not re-verified.
23 juil. 2026
conditional -> conditional (rationale updated): On July 22, 2026, White House OSTP Director Michael Kratsios publicly accused Moonshot AI of industrial-scale distillation of Anthropic's Fable 5 to train Kimi K3. Treasury Secretary Bessent threatened sanctions and Entity List designations.

Sources

Journal de vérification

  • Tarifs· Aucun changement

    Agent automatisé

  • Tarifs· Aucun changement

    Agent automatisé

  • Contenu· Mis à jour

    Agent automatisé

    Price $3/$15 and 1M context unchanged. Summary was 28 words framing a live investigation with no product context; rewritten (verdict_set, conditional held) from stored sourced children, with the July 22 dispute attributed and flagged as not re-verified this run.

  • Profil· Mis à jour

    Agent automatisé

    Pricing matches the Kimi platform rate card: $3.00 input, $0.30 cached input, $15.00 output per 1M, and a 1,048,576-token context. The children contradicted the entity's own pricing_context ('Weights live (Jul 27)'): an avoid_for row said 'weights not yet downloadable' and a not_for_whom row said 'wait for Jul 27'. Hugging Face moonshotai/Kimi-K3 serves the weights under the Kimi K3 License, and both rows were rewritten via model_update. pricing_url moved from the blog to the platform rate card (platform.moonshot.ai 301-redirects to platform.kimi.ai), and the rate card and HF page were added as sources. Scalars unchanged. No verdict change. Flag: the summary is about 28 words, below the 55-90 word standard, and the sanctions threat from July has not been re-checked.

  • Tarifs· Aucun changement

    Agent automatisé

  • Tarifs· Mis à jour

    Agent automatisé

    Kimi K3 open weights went live on Hugging Face Jul 26 evening EDT (one day ahead of Jul 27 schedule). Model is now truly open-weight under Modified MIT license. Pricing unchanged at $3/$15. This is a material status change from "weights not yet downloadable" to "weights live on Hugging Face".

  • Tarifs· Aucun changement

    Agent automatisé

Comment nous évaluons