Kimi K3
Run frontier-level coding and reasoning tasks on a 2.8-trillion-parameter (104B active) open-weight model. Full weights are released under the Kimi K3 License, so self-hosting is available now, not just on the hosted API.
The verdict on Kimi K3: Developers who need long-horizon reasoning and multi-file coding work from an open-weight model Kimi K3 is aimed at developers who need long-horizon reasoning and multi-file coding work from an open-weight model that competes with closed frontier performance on benchmarks. Pricing: Pay-per-token API ($0.30-$15 per million tokens) / full weights released under the Kimi K3 License, free to download and self-host (compute costs apply).
Best For
Developers who need long-horizon reasoning and multi-file coding work from an open-weight model
Standout Feature
A 2.8-trillion-parameter model competitive with closed frontier models on independent coding and math benchmarks
TL;DR
Genuinely impressive open-weight performance, self-hosting the full model requires serious GPU infrastructure and it's very new.
Alternatives
Fit Score: use-case reference
How this is scoredKimi K3 scored 1 to 5 on five dimensions for the named use cases below. Part of our catalog-wide use-case reference, not a product-quality review rating.
Open-weight long-horizon reasoning
3.4/5- Capability
- 4/5
- Reliability
- 3/5
- Value
- 3/5
- Ease
- 3/5
- Compatibility
- 4/5
Run multi-step reasoning with tool-calling support and an open-weight deployment option. The release has limited long-term reliability data, so assess it on the actual workflow rather than assuming benchmark claims guarantee results. API billing is usage-based, while self-hosting requires substantial GPU capacity.
Open-weight multi-file coding
3.4/5- Capability
- 4/5
- Reliability
- 3/5
- Value
- 3/5
- Ease
- 3/5
- Compatibility
- 4/5
Work across multiple code files through an API or a self-hosted open-weight deployment. Limited long-term reliability evidence makes project-level validation important before relying on generated changes. Usage-based API fees and substantial self-hosting hardware needs mean the free weight download is not a free coding service.
Editorial judgment grounded in this tool's own published pricing, features, and our existing pros and cons, not a certification or a measured benchmark. Product quality is assessed separately under our review methodology.
Overview
Not yet reviewedProduct quality is assessed from original evidence. How reviews work.
Kimi K3 is Moonshot AI's open-weight model, built for long-horizon reasoning and coding work that runs across many steps instead of a single response. It's a mixture-of-experts model with 2.8 trillion total parameters and 104 billion active parameters, with roughly 1 million tokens of context per Moonshot's own model card. It handles multi-file codebases, agentic tool-calling loops, and extended chains of reasoning at a scale that puts it in range of the top closed frontier models on independent coding and math benchmarks.
Access is available through the API at usage-based rates, and the full model weights are now released under the Kimi K3 License with deployment guidance on Moonshot's Hugging Face repository, so teams can self-host today rather than depend on a single vendor's uptime and pricing. Downloading the weights themselves carries no charge, but running a model this size takes serious infrastructure, so most users will start on the hosted API and only move to self-hosting once they have the hardware to justify the nonzero compute cost. Kimi K3 fits engineering teams and researchers who want frontier-level reasoning and coding performance with the option to run the model on their own infrastructure instead of staying locked into a single API.
The source of the legacy catalog rating and review count has not been substantiated. Neither is presented here as a verified review rating or included in review structured data.
Workflow Fit Scores are separate, use-case-specific editorial judgments, not user-review averages. Read the Fit Score methodology.
- • Competitive with closed frontier models on coding and math benchmarks
- • Open weights mean no long-term vendor lock-in
- • Pay-per-token API access available immediately, no waitlist
- • Self-hosting the full model requires substantial GPU infrastructure
- • Very new release with limited independent long-term reliability data
Key Features
- 2.8-trillion-parameter (104B active) mixture-of-experts architecture, ~1M-token context
- Long-horizon multi-step reasoning and agentic tool-calling
- Multi-file codebase support for real-world coding tasks
- Self-hostable weights released under the Kimi K3 License (no charge to download; running them takes serious GPU infrastructure)
Trust & Data
Verified 2026-09
- Training on your data
- Opt-out availableThe access route matters a great deal here: Kimi's API help page and business terms state API and enterprise input and output are never used for model training, but the consumer web-chat terms, the actual kimi.com product cataloged here, state that prompts and generated content are used to train and optimize models by default, with opt-out only available by emailing support manually, a 5 to 7 business day process with no in-product toggle. The general terms and the narrower help-page and enterprise assurances genuinely conflict, so casual web users get materially weaker protection than API or enterprise customers.
- Compliance
- Not published
- Retention
- Not published
- Export
- Partial export
Sources: kimi.com/userprivacy, kimi.com/data usage
As published by the vendor. Verify independently before purchase decisions.
Our Take
At 2.8 trillion parameters, agentic tool-calling loops and extended reasoning chains are practical rather than theoretical. Open weights mean no long-term vendor lock-in, and the option to self-host matters for teams with data sovereignty requirements. Two honest constraints apply: self-hosting the full model requires substantial GPU infrastructure, and it's a new release with limited long-term reliability data. If coding and reasoning at frontier scale is the requirement, evaluate it now, but plan for real infrastructure costs.
From the Blog
Coverage and comparisons that mention Kimi K3.