AI Model Radar
中文 EN

Kimi K3

Moonshot AI / Kimi
Model overview

kimi-k3

Moonshot/Kimi's July 2026 native-multimodal model for software engineering, knowledge work, deep reasoning, and long-horizon agent tasks.

general LLMreasoningcodingmultimodalvision-languageagentlong contextchatcode generationsoftware engineering agentdata analysisdocument understandingenterprise access
External observationsCross-model riskCompareCompanyRankingsPublic signals
Model risk

Current observations

2current adverse observations
Negative / riskModel behaviorRuntime behaviorOfficial · Current risk

Incomplete historical thinking or mid-session model switching can destabilize output

Kimi requires complete historical thinking content to be passed back and warns that switching from another model to K3 mid-session can materially reduce generation stability.

Kimi K3: Open Frontier IntelligenceHigh confidenceObserved 2026-07-16Source as of 2026-07-20 Action

Pin the model ID before migration and validate preservation and replay of thinking history in real multi-turn tool workflows.

Open source
Negative / riskModel behaviorRuntime behaviorOfficial · Current risk

Official guidance acknowledges excessive proactiveness on ambiguous tasks

Kimi's release guidance says K3 can sometimes take more action than needed for ambiguous intent or simple questions.

Kimi K3: Open Frontier IntelligenceHigh confidenceObserved 2026-07-16Source as of 2026-07-20 Action

Production agents should constrain tool permissions, define stop conditions, and validate authorization boundaries with real tasks.

Open source
Historical records 1
HistoricalAPI / serviceAccess restrictionsOfficial · Historical

Full weights were published on July 27, closing the earlier availability gap

Moonshot has published the full Kimi K3 weights, model card, and Kimi K3 License; the earlier unavailability observation is no longer a current risk.

MoonshotAI/Kimi-K3 official repositoryHigh confidenceObserved 2026-07-16Source as of 2026-08-04Ended 2026-07-27 Action

Local deployment still requires review of the Kimi K3 License; the open weights do not include Kimi.com's hosted agent orchestration, tool routing, or serving stack.

Open source
Representative version

kimi-k3

CompanyMoonshot AI / Kimi
Release2026-07-16
Release typeModel release
Parameters2,800B / 104B active
ArchitectureKimi Delta Attention + Attention Residuals + Stable LatentMoE
Context1,048,576
Input modalitiestext / image / video
Output modalitiestext / code / tool_call
Tool use / function callingConfirmed: Tool use, Function calling
Structured output / reasoningConfirmed: Structured output, Reasoning
Vision / audio / videoConfirmed: Vision, Video
Open weights / LicenseWeights available / Kimi K3 License
License sourcemoonshot-kimi-kimi-k3-weights-2026-07-27 · 2026-08-04
Vendor APIAvailable
Kimi K3: Open Frontier Intelligencemoonshot-kimi-kimi-k3-current · 2026-07-16
OpenRouterOpenRouter listed
PricingOfficial · 2026-07-16: cache-miss input $3 / output $15 · cache-hit input $0.30
OpenRouter · 2026-08-11: $3 / $15
input / output, per 1M tokens
BenchmarkNo reliable public source
Public signals2 community discussions / 1 activity snapshot
Use casesSoftware engineering agents / Repository-scale coding / Knowledge work / Long-horizon reasoning
Claim evidenceRelease eventKimi K3: Open Frontier Intelligencekimi-k3 · 2026-07-16
Model specifications / Weight checkpointmoonshotai/Kimi-K3 official model card and weightsKimi K3 · Kimi-K3 full weights · Kimi-K3 full weights
OpenRouter context / capabilities / pricingOpenRouter public models APImoonshotai/kimi-k3 · 2026-08-11T06:28:09.895Z
OpenRouter

Pricing / context / availability

OpenRouterOpenRouter listed
OpenRouter listingmoonshotai/kimi-k3
Mapped versionkimi-k3
moonshot-kimi-kimi-k3-current
PricingOfficial · 2026-07-16: cache-miss input $3 / output $15 · cache-hit input $0.30
OpenRouter · 2026-08-11: $3 / $15
input / output, per 1M tokens
Context1048576
Speed/LatencyNo reliable public source
Snapshot time2026-08-11T06:28:09.895Z
SourceView source
Source-bounded public ranking

OpenRouter weekly token-usage ranking

SourceModel / versionMetricRankValueCheckedLink
OpenRouterMoonshotAI: Kimi K3 0715
moonshot-kimi-kimi-k3-current		Weekly token usage on OpenRouter public rankings API121,420,233,247,221 tokens2026-08-11View source

OpenRouter weekly prompt-plus-completion token usage only; this is not a model-quality or whole-market ranking.

Verified limitations

Usage boundaries

  • The API supports low, high, and max reasoning effort, with max as the default; disabling thinking falls back to Kimi K2.6.
  • Full model weights were published on 2026-07-27 under the Kimi K3 License; the weights do not include Kimi.com's hosted agent orchestration, tool routing, or serving stack.
Public activity counters

Hugging Face model-card activity snapshot

PlatformEntryLast-month downloadsCurrent likesCurrent SpacesTimeSource
huggingface
Current HF activity snapshot		moonshotai/Kimi-K31,510,03210,47932View source

Downloads are Hugging Face's trailing-month counter; likes and Spaces are snapshots at collection time.

Vertical comparison

Current available versions

VersionReleaseParametersContextInputOutputOpen weights / LicenseVendor APISource tier
Kimi-K3 full weights2026-07-272,800B / 104B active1,048,576text / image / videotext / code / tool_callWeights available / Kimi K3 LicenseAvailableOfficial source
kimi-k32026-07-162,800B total / active parameters not disclosed1,048,576text / image / videotext / code / tool_callNo open weights currently / weight license not disclosedAvailableOfficial source
Missing public information

Public information not yet available

No reliable benchmark sourceNo linkable model-specific reviewSpeed/latency not disclosed