UTC

Kimi K2 Thinking

Moonshot AI · open weights · served by 15 providers · tool calling, reasoning

moonshotai/kimi-k2-thinking

Weights
Open
Released
2025-11-06
Knowledge cutoff
2024-08
Context window
262K
Deprecation dates

When each platform deprecates Kimi K2 Thinking, as LiteLLM records it. Plan the move before the date: after it, the platform may stop serving the model or stop updating it.

  • Amazon Bedrockdeprecates 2027-03-30moonshot.kimi-k2-thinkingsuccessor: moonshotai.kimi-k3.1
  • Azure AI Foundrydeprecated 2026-03-29azure_ai/kimi-k2-thinkingsuccessor: azure_ai/kimi-k2.5
  • Basetendeprecated 2026-10-16baseten/moonshotai/Kimi-K2-Thinkingsuccessor: baseten/moonshotai/Kimi-K3.1-Pro
Every upcoming deprecation →
Hugging Face repository

moonshotai/Kimi-K2-Thinking at commit a51ccc050d, last changed 2026-01-30

Licence
modified-mit
Access
Open download
Library
transformers
Formats
.safetensors
  • Loading it runs code from the repository (trust_remote_code). Read that code, and pin the commit.
  • No model signature file is published with the weights.

What the Hub's scanners report

35 weight files, 35 with a scan result. No scanner reports an issue with them.

The model at this commit, its 35 weight files with their SHA-256.

Facts from the Hugging Face Hub and its file scanners (Protect AI, Cisco, Hugging Face picklescan, VirusTotal, JFrog), as of 2026-11-18. ShieldMap does not copy model cards.

Who serves it
  • 302.AIkimi-k2-thinking
  • AI/ML APImoonshotai/kimi-k2-thinking
  • Amazon Bedrockmoonshot.kimi-k2-thinkingdeprecated
  • Eden AIamazon/moonshot.kimi-k2-thinking
  • Heliconekimi-k2-thinking
  • Hugging Facemoonshotai/Kimi-K2-Thinking
  • IO.NETmoonshotai/Kimi-K2-Thinking
  • Kilo Gatewaymoonshotai/kimi-k2-thinking
  • Meganovamoonshotai/Kimi-K2-Thinking
  • NanoGPTmoonshotai/kimi-k2-thinking
  • Novita AImoonshotai/kimi-k2-thinking
  • OpenCodekimi-k2-thinkingdeprecated
  • OpenRoutermoonshotai/kimi-k2-thinking
  • Poenovita/kimi-k2-thinkingdeprecated
  • Qiniumoonshotai/kimi-k2-thinkingdeprecated
Running it yourself

Open weights are loaded and served by code that has its own advisories. These inference servers and model loaders are on ShieldMap; load weights only from the publisher's own repository, and prefer safetensors over formats that can run code.

How supply-chain attacks on models work →

Sources: models.dev (MIT) and the LiteLLM model catalogue (MIT). Deprecation dates are as each platform published them and as LiteLLM records them; check the platform before you rely on one. Data as of 2026-11-22.