Model AtlasA selection map of agent foundation models 中文
Start researchModelsPapersGuideCompare 中文

Search the atlas

Esc to close · Cmd/Ctrl + K to open
Moonshot AI

Kimi K3

Instruction-tuned · 2026-07-16 · Verified
Research summary

Research summary

Good for
Inference, LoRA, SFT, RL
Not suitable / limits
No explicit limitation evidence
How to get it
Open weights
Hardware tier
Inference: Multi-GPU · LoRA: Multi-GPU
Paper adoption
No paper-adoption record
Reproduction risk
Some key facts need verification
Evidence state
Verified
Model positioningGeneralReasoningCodingAgentTool useMultimodal

Access and distribution ladder

ProductYesavailable
APIYesavailable
WeightsYesYes
BasePending verificationNot reported
Finetuning allowedYesYes
Derivative distributionPending verificationNot reported

At a glance

Model positioning
General, Reasoning, Coding, Agent, Tool use, Multimodal
Architecture
Mixture of experts (MoE)
Parameters
Not reported total / Not reported active
Active parameters
Not reported
Experts
896
Experts / token
16
Context length
Not reported
Modalities
Text, Image
Base checkpoint
Not reported
Finetuning allowed
Yes
License
Kimi K3 License

Unresolved fields

These fields stay explicit instead of being guessed; verify them against the source or model card.

10 key fields need verification

How to read unknown states

These states intentionally preserve evidence boundaries; they are not false, zero, or definite facts.

Not disclosedThe official source does not disclose this; do not infer that it is absent.
The official source does not disclose this; do not infer that it is absent.
Not applicableThis field does not apply to this record; it is not missing or negative.
This field does not apply to this record; it is not missing or negative.
Not reportedThe source or paper did not report it; this is not false, zero, or confirmed absence.
The source or paper did not report it; this is not false, zero, or confirmed absence.
Not verifiedThe field is retained, but current evidence is not verified; add first-party evidence before relying on it.
The field is retained, but current evidence is not verified; add first-party evidence before relying on it.
Conflicting evidenceFirst-party sources conflict; no single value is selected.
First-party sources conflict; no single value is selected.
Not publishedThe record is not published yet; this is not the same as a broken source link.
The record is not published yet; this is not the same as a broken source link.
UnavailableThe current source is unavailable; do not infer that the fact does not exist.
The current source is unavailable; do not infer that the fact does not exist.

Access & distribution

Weights status
Released
Full weights date
2026-07-27
API status
Available
API model IDs
kimi-k3
Product status
Available
Product surfaces
Kimi App, Kimi Agent, Kimi Code
Openness classification
Open weight
Custom license
Yes
Derivative distribution
Conditional
Commercial use
Conditional
License conditions
Redistribution and commercial deployment are subject to the Kimi K3 License.; This summary is not legal advice.
Model revision
Pin the model revision
Preserve reasoning history
Yes
Reproducibility notes
Preserve complete reasoning and tool-call history across multi-turn tool use.

Specializations & research readiness

GeneralReasoningCodingAgentTool useMultimodal
Inference
Yes
LoRA
Yes
SFT
Yes
RL
Yes
Transformers
Not reported
vLLM
Yes
SGLang
Yes
veRL recipe
Not reported

Hardware tiers

Inference
Multi-GPU
LoRA
Multi-GPU
Full-parameter SFT
Multi-GPU
RL
Not reported

Paper adoption

No paper-adoption records yet. No record does not mean no usage.

Replacement candidates and family versions

Candidates are derived from catalog family, generation, and release metadata; this is not a quality score or a substitute for experiments.