Chinese AI Tools
ProductsModelsIntegrationsRankingsLatest changes
Availability TrackerUse casesSubmit a toolAccount
ZHSearch

Chinese AI Tools

Independent directory for Chinese AI products. Product availability, pricing and terms can change. Verify before commercial use.

Editorial standardsClaim productUpdate infoGet featuredAdvertise

Moonshot AI

Kimi K3

Kimi K3 is Moonshot AI's 2.8-trillion-parameter Mixture-of-Experts flagship built on Kimi Delta Attention and Attention Residuals, with native vision and a 1M-token context window. It effectively activates 16 of 896 experts and targets long-horizon coding, agentic knowledge work and reasoning. K3 is available in Kimi, Kimi Work, Kimi Code and the Kimi API as kimi-k3. The launch API price is $0.30 per million cache-hit input tokens, $3 per million cache-miss input tokens and $15 per million output tokens. Moonshot released the full Kimi K3 weights, model code and custom license on Hugging Face on July 27, 2026; the repository contains 96 weight shards and an MXFP4-compressed checkpoint. The official limitations still warn that harnesses which fail to preserve K3's full thinking history can produce unstable quality, and recommend a verified-compatible harness such as Kimi Code.

Globally availableFull English UIPublic APIPaid

Editorial verdict

Best for

Teams evaluating a frontier-scale Chinese model for repository-scale coding, multimodal technical work, agentic research and long-context knowledge workflows.

Avoid if

Avoid it when you lack the infrastructure for a 2.8T model, need a low-cost short-context model, or use a harness that cannot reliably retain full reasoning history.

Why it matters

Kimi K3 merits a separate LLM & API profile because it is a flagship model with its own architecture, context length, hosted pricing and public open-weight release, rather than a minor K2.7 Code revision.

Trust: 4/4 sources verified, recently checkedCoverage: 100/100

Pricing

$0.30/M cache-hit input, $3/M cache-miss input and $15/M output tokens

Payment

Kimi API pay-as-you-go billing, Kimi membership, Kimi Work subscription, Enterprise sales

Commercial use

Commercial use should follow the current product, API, model license and billing terms.

Privacy

Review prompt, file, media upload, retention and training-use terms before sensitive workloads.

Use-case fit

Long-horizon coding agents

Strong

Use Kimi Code or a verified-compatible harness for repository navigation, terminal-tool orchestration and extended engineering sessions.

Multimodal software and technical work

Strong

Apply native vision to screenshot-driven frontend, game, CAD and visual-debugging workflows alongside code generation.

Agentic research and knowledge work

Strong

Evaluate K3 in Kimi Work or the API for deep research, interactive reports, dashboards and multi-step analysis.

Global user checklist

RegistrationConfirmedK3 is publicly available through Kimi, Kimi Work, Kimi Code and the Kimi API platform.
English UIConfirmedThe announcement and linked Kimi product surfaces are English-facing.
API and docsPartialThe launch page documents the kimi-k3 model ID and pricing, while detailed API parameters and compatibility guidance should be rechecked in the platform docs.
International paymentConfirmedThe official launch page lists $0.30/M cache-hit input, $3/M cache-miss input and $15/M output token pricing.
Commercial usePartialHosted use follows Kimi API, membership or enterprise terms; self-hosting follows the separate Kimi K3 model license, which must be reviewed for the intended deployment.
Data and privacy termsPartialReview the appropriate Kimi consumer, API or enterprise privacy terms before submitting repositories, media or sensitive research data.

Hosted access and API pricing were rechecked on August 4, 2026, and the July 27 open-weight release is now linked. Recheck thinking-effort controls, capacity and license suitability before production deployment.

Pros

  • - 1M-token context with native image and video understanding
  • - Available now across Kimi, Kimi Work, Kimi Code and the Kimi API
  • - Designed for long-horizon coding, terminal tools, visual software work and agentic knowledge workflows
  • - Published API pricing separates cache-hit from cache-miss input

Cons

  • - Full weights are public, but the 2.8T model and roughly 1.56 TB repository require specialized multi-accelerator infrastructure
  • - Launch behavior is fixed to max thinking effort; lower effort modes are not yet available
  • - Reasoning quality may become unstable when prior thinking history is dropped or a session switches from another model
  • - Benchmark results and long autonomous case studies are vendor-reported and use varied agent harnesses

Decision paths

kimi-k2-7-code

minimax-m3

zhipu-glm

qwen

deepseek

Sources

Kimi K3 official announcement

official · en · verified 2026-08-04

Confirms K3 architecture, 2.8T parameters, 1M context, native vision, hosted availability, model ID, pricing, benchmark disclosures and limitations.

Kimi K3 model repository

repository · en · verified 2026-08-04

Confirms the public July 27 weight release, custom Kimi K3 license, model code, MXFP4 checkpoint and 96 weight shards.

Kimi API platform

docs · en · verified 2026-07-17

Official platform entry linked from the K3 availability section for API access.

Kimi Code

official · en · verified 2026-07-17

Official Kimi Code surface linked as a K3 access path and recommended-compatible harness.

Last checked: 2026-08-04

Reviews

Availability snapshot

Availability
available
English UI
full
API
available
Rating
4.5 (0)

Latest updates

Latest changes
Open source · 2026-08-04

Kimi K3 full weights are now public

Moonshot AI published Kimi K3's full weights, model code and custom model license on Hugging Face on July 27. The public repository contains an MXFP4-compressed checkpoint split across 96 weight shards, closing the launch announcement's open-weight commitment while leaving substantial infrastructure and license-review requirements for self-hosting.

Release · 2026-07-17

Moonshot launches Kimi K3

Moonshot AI launched Kimi K3 as a hosted 2.8T-parameter, native-vision MoE model with a 1M-token context window. It became available in Kimi, Kimi Work, Kimi Code and the Kimi API as kimi-k3 at $0.30/M cache-hit input tokens, $3/M cache-miss input tokens and $15/M output tokens. At launch, Moonshot scheduled the full-weight release for July 27; that release is now tracked separately. Published limitations include sensitivity to dropped thinking history and possible over-proactiveness on ambiguous tasks.

Submit a review