Kimi K3
Moonshot’s 2.8T MoE (104B active) with 1M context. Kimi K3 weights are the downloadable checkpoint. Kimi K3.1 Pro (2026-10-23) is the current API flagship; Moonshot says those weights will be released by 2026-11-04.
Moonshot’s Kimi K3 family — 2.8T MoE (104B active), 1M context, native vision. Hugging Face weights under the Kimi K3 License (moonshotai/Kimi-K3) remain the downloadable checkpoint, including in GitHub Copilot. Kimi K3.1 Pro (2026-10-23) is a post-trained K3 and the current API flagship: kimi-k3.1-pro, plus kimi-k3.1-pro-highspeed, which Moonshot says is 2–3× faster. Both are on Kimi.ai, Kimi Work, Kimi Code, and the Kimi API. Moonshot says full K3.1 Pro weights will be released by 2026-11-04. Kimi Enterprise access to K3.1 Pro is off by default. The launch post does not publish token prices.
Why Kimi K3 matters
Kimi K3 pushed open weights into the 3T-class MoE tier with competitive long-horizon coding and native vision. K3 weights under the Kimi K3 License remain what you can self-host, including in GitHub Copilot. Kimi K3.1 Pro (kimi-k3.1-pro, plus the Highspeed id kimi-k3.1-pro-highspeed) is the current API post-train on Kimi.ai, Kimi Work, Kimi Code, and the Kimi API.
Vision · Tool calling · Thinking · MCP · Coding
Last reviewed: 26 October 2026
When to choose Kimi K3
Decision guidance for architects—not a feature list.
Best for
- Long-horizon open coding agents
- Open multimodal research and products
- 1M-context knowledge work
- Self-host frontier open deployments
Avoid if
- You need MIT/Apache-simple licensing for large-scale MaaS
- You cannot operate very large MoE serving footprints
- You need K3.1 Pro weights before Moonshot’s 2026-11-04 release
Strengths
DataAIHub qualitative scores from 1 to 5, for architects. Not official benchmark results or a public ranking.
- Open frontier scale★★★★★
- Agentic coding★★★★★
- Multimodal + long context★★★★★
- License simplicity★★☆☆☆
- Serving ease★★☆☆☆
Ecosystem
Built by
- Moonshot AI
Kimi K3 is Moonshot’s open 2.8T MoE. Kimi K3.1 Pro (2026-10-23) is the current API flagship; Moonshot says those weights will be released by 2026-11-04.
Competes with
- Llama 4
Competes as an open generalist when frontier scale and multimodality matter.
- DeepSeek V4
Peer open MoE for agentic coding and 1M-context deployments.
- GPT-5.6
Open weights alternative for long-horizon coding and knowledge work.
Works with
Recommended for
- Large language models
Reference open 3T-class MoE in today’s LLM landscape.
- Llama Models
Compare open-weight deployability and licensing vs Llama stacks.
How Kimi K3 evolved
Key moments in chronological order.
- Model
Kimi K3.1 Pro, a post-trained K3, is the current API flagship (kimi-k3.1-pro). Kimi K3.1 Pro Highspeed (kimi-k3.1-pro-highspeed) is the faster API variant; Moonshot says it runs 2–3× the speed of K3.1 Pro. Both are on Kimi.ai, Kimi Work, Kimi Code, and the Kimi API, with Zero Data Retention for API customers. Enterprise access is off by default. Moonshot says full weights will be released by 2026-11-04, and that the open-weight release removes cyber-related mid-training data and RL environments. On Moonshot’s internal xdev-bench-hard, pass@3 is 81.0% for the released checkpoint and 90.5% for the unfiltered pipeline. Hugging Face Kimi K3 weights remain the downloadable checkpoint.
- Platform
Kimi K3 rolls out as a Copilot model option on Pro, Pro+, Max, Business, and Enterprise plans.
- Platform
Kimi K3 on Databricks Unity AI Gateway
Databricks hosts Kimi K3 via Foundation Model API with Unity AI Gateway governance (US hosting; AWS and GCP workspaces).
- Open source
Open weights on Hugging Face
Full 2.8T checkpoint published under the Kimi K3 License (not MIT/Apache).
- API
Kimi K3 API / product launch
Moonshot launches Kimi K3 via platform.kimi.ai and consumer Kimi surfaces.
- Research
KDA + Stable LatentMoE
Kimi Delta Attention and 16-of-896 expert MoE target long-context efficiency.
- Model
Native multimodal + 1M context
Text, image, and video in one open model with a 1,048,576-token window.
Overview
Capabilities
- Vision: Yes
- Audio input: No
- Tool calling: Yes
- Thinking: Yes
- MCP: Yes
- Coding: Yes
- Structured output: Yes
Technical specifications
- Provider
- Moonshot AI
- License
- Kimi K3 License (custom; commercial scale conditions)
- Context window
- 1.0M
- Parameters
- MoE 2.8T total / 104B activated
- Architecture
- Stable LatentMoE (KDA + Attention Residuals)
- Release
- 2026-07
- Modalities
- Text, Image, Video
- Vision
- Yes
- Audio input
- No
- Tool calling
- Yes
- Thinking
- Yes
- MCP
- Yes
- Open weights
- Yes
- API
- Yes
- Pricing (input)
- Moonshot / Kimi API (verify platform.kimi.ai)
- Pricing (output)
- Moonshot / Kimi API (verify platform.kimi.ai)
Supported modalities
Text · Image · Video
Context window
1.0M (1,048,576 tokens)
Pricing
K3 weights can be self-hosted; large-scale MaaS may need a separate Moonshot agreement. The K3.1 Pro launch post does not list token prices — confirm on platform.kimi.ai.
Availability
Downloadable weights are Kimi K3 (API id kimi-k3) at huggingface.co/moonshotai/Kimi-K3, including GitHub Copilot (Pro through Enterprise). Current API flagship ids are kimi-k3.1-pro and kimi-k3.1-pro-highspeed. K3.1 Pro weights are scheduled for 2026-11-04. Zero Data Retention is available for Kimi API customers.
Use cases
- Long-horizon agentic coding
- Open multimodal research and products
- 1M-context knowledge work
- Self-hosted frontier open deployments
Strengths
- 2.8T open MoE with 1M context
- Native multimodal + 1M context
- Competitive agentic coding vs closed frontier peers
Limitations
- Custom license — not MIT/Apache; check commercial clauses
- Very large download / serving footprint
- Ecosystem younger than Llama/Qwen stacks
- K3.1 Pro weights are not public yet (Moonshot: by 2026-11-04)
- Moonshot says the K3.1 Pro open-weight release removes cyber-related mid-training data and RL environments
Related guides
Related benchmarks
Related research
- Introducing Kimi K3.1 Pro
Original Paper
- Kimi K3 model card (Hugging Face)
Original Paper
- Top Open-Source & Open-Weight AI Models 2026
Evaluation
Related GitHub
Related tools
Related rankings
Companies
Explore more models
- DeepSeek V4DeepSeekDeepSeek’s V4 generation — live Flash is V4.1-Flash-1017 (API id deepseek-flash, upgraded 2026-10-17): Causal Encoder–Decoder MoE, 552B backbone (8B active prefill / 16B decode), native image+text, and 1M context. API id and prices are unchanged from the 2026-09-10 Flash list. MIT weights on Hugging Face. Legacy deepseek-v4-flash and deepseek-v4-flash-vision-exp aliases temporarily route here. deepseek-v4-pro still serves V4-Pro-0813 after 2026-09-14 at unchanged Pro rates (DeepSeek withdrew the planned Flash reroute).
- MiMo V2.6XiaomiXiaomi’s MiMo-V2.6 family — MIT-licensed native-omnimodal MoE with 1M context. Flagship open checkpoint is MiMo-V2.6-Pro-RL (1.02T total / 42B activated); Flash-RL is 309B / 15B activated. Hosted API ids: mimo-v2.6-pro, mimo-v2.6-flash, mimo-v2.6-pro-ultraspeed. Xiaomi publishes an Artificial Analysis Intelligence Index score of 46 for Pro; that figure is vendor-published and is not a DataAIHub ranking.
- Muse GlimmerMetaMeta Superintelligence Labs’ Muse Glimmer — Apache-2.0 ~30B dense multimodal agent model for on-device and single-GPU local agents. Sibling to closed Muse Spark and Muse Wonder; distinct from Llama 4.
- Qwen3AlibabaAlibaba’s Qwen3 family spanning Qwen3.8-Max (2.4T MoE / 95B active; live cloud alias qwen3.8-max routes to the 0902 snapshot as of 2026-09-05), open Qwen3.8-27B (dense VLM, Apache-2.0), Qwen3.8-Flash-Next (125B / 6B active multimodal MoE + 51B n-gram embeddings)—a Qwen4 architecture preview—and Qwen3.8-Omni-Flash (hosted 2026-09-17: text/image/audio/video in, text out, 1M context; API id qwen3.8-omni-flash). Production Qwen3.8-Flash on QwenCloud adds 1M-default context and built-in tools atop the Flash-Next design.
- Llama 4MetaMeta’s Llama 4 family — open-weight multimodal models designed for research and commercial use under Meta’s community license.
- Claude FableAnthropicAnthropic’s Claude Fable 5.1 — the most capable widely released Claude for long-horizon agents, deep reasoning, and demanding coding workflows. Mythos 5.1 is the same underlying model with relaxed cyber/life-sciences safeguards for trusted-access programs.