Moonshot AI Coding and Agent Model

Kimi K3 Model Profile

Kimi K3 is Moonshot AI's 2.8T-parameter model for coding, agent workflows, long-context reasoning, and visual understanding. It is most relevant when you need large-repo coding, 1M-token context, or an API model to compare through OpenRouter.

Updated July 23, 2026Release date: July 16, 2026Model ID: k3
Use Kimi K3 if

You need coding-agent work, long-context analysis, repository reasoning, or a model to test against OpenRouter alternatives.

Be careful if

You need stable low-cost production traffic, guaranteed subscription capacity, or a clear self-hosting hardware plan today.

Best first test

Run the same coding, document, and agent prompts against Kimi K3, your current model, and one cheaper fallback model.

Kimi K3 Facts at a Glance

Kimi K3 is a Moonshot AI model released on July 16, 2026. Its two headline numbers are 2.8T total parameters and up to a 1M-token context window, with actual access depending on the provider and membership tier.

ItemCurrent fact
OwnerMoonshot AI, the company behind Kimi
Release dateJuly 16, 2026
Model size2.8T total parameters
Context windowUp to 1M tokens, depending on access tier
Official Kimi Code IDk3
Reasoning controlsLow, high, and max thinking levels
ModalityText, code, agents, and native visual understanding
OpenRouter listingmoonshotai/kimi-k3

What Kimi K3 Is Best For

Kimi K3 is strongest when the input is large or the workflow has many steps. For a simple short chat, it may be more model than you need; for a large repository, long spec, or coding agent flow, it is worth testing.

Use caseFitWhy it matters
Large-repo codingHighKimi positions K3 for coding and agent work, with long-context support for large codebases.
Agent workflowsHighUse K3 when the task requires planning, tool use, code edits, or multi-step execution.
Long-document analysisHighThe 1M-token context is the main reason to test K3 for documents, logs, specs, and research packs.
Visual understandingMedium to highK3 adds native visual understanding, but verify output quality with your own task set.
Low-cost short chatSituationalA smaller or faster model may be cheaper when the task does not need K3 reasoning or context.

Kimi K3 Release Date and Timeline

DateUpdate
July 16, 2026Kimi K3 was released and made available in Kimi Code.
July 20, 2026Kimi Code v0.28 release notes expanded K3 documentation and membership notes.
July 20, 2026AP reported that Moonshot AI temporarily paused new subscriptions because demand exceeded capacity.
July 22, 2026Kimi Code v0.29 added clearer K3 session and model-switching notes.

Kimi K3 Access, Subscription, and OpenRouter

The fastest path depends on your job. Use the Kimi app for general testing, Kimi Code for developer workflows, and OpenRouter when you want one API layer for side-by-side model comparison.

RouteBest forAccess note
Kimi appBest for chat, long documents, research, and everyday assistant use.Available through Kimi web and mobile apps.
Kimi CodeBest for coding agents, large repositories, code review, and multi-file edits.K3 access starts at Moderato; 1M context starts at Allegretto and above.
Kimi APIBest for developers who want to connect K3 through API-compatible clients.Use the official Kimi coding API endpoint and model ID k3 where supported.
OpenRouterBest for testing K3 next to other models in one API.OpenRouter lists Kimi K3 with 1.05M context and per-token pricing.

Kimi K3 Pricing and Limits

Pricing and capacity can change quickly for a newly released model. Treat this as a dated snapshot and verify the provider page before buying a plan or building production usage around it.

Access typePrice or limitImportant caveat
Kimi subscriptionPlan-basedK3 access and context limits depend on Kimi membership tier.
OpenRouter$3 per 1M input tokens; $15 per 1M output tokensShown on OpenRouter for moonshotai/kimi-k3 as of this update.
Self-hostingNot enough official serving detail for a safe estimateDo not infer hardware cost from 2.8T parameters alone. Wait for model card and deployment guidance.

Kimi K3 Benchmarks and Comparisons

Kimi K3 should be compared by task, not by one headline score. For coding, test multi-file edits and bug fixes; for long context, test whether it retrieves the right detail near the middle of a long document; for agents, test tool-call reliability and recovery after mistakes.

ComparisonWhat to comparePractical recommendation
Kimi K3 vs Kimi K2.7 CodeK3 is the newer 2.8T model with 1M-context access and reasoning-level controls.Use K3 for the newest coding and agent tasks; use K2.7 only when your workflow already depends on it.
Kimi K3 vs GLM 5.2Compare by coding benchmark, context limit, Chinese/English quality, API price, and deployment options.Build a task-based test set before declaring a winner.
Kimi K3 vs Fable 5Treat this as an open comparison until Fable 5 release, pricing, and benchmark details are confirmed.Avoid winner claims until both models are tested on the same tasks.
Kimi K3 vs Opus 4.8Compare coding agent reliability, long-context retrieval, tool use, speed, and price.Use side-by-side prompts instead of generic benchmark screenshots.

Hardware Requirements and Open-Weight Caveat

Kimi documentation says K3 was open-sourced, and OpenRouter describes it as open-weight. That does not automatically mean a simple local setup. For Kimi K3 hardware requirements, use the official model card, weights license, quantization notes, and serving benchmarks before estimating GPUs, memory, throughput, or total hosting cost.

How to Use Kimi K3

  1. 1. Start in Kimi app: open Kimi for chat, document analysis, and everyday model testing.
  2. 2. Use Kimi Code for coding: check the Kimi Code model docs and select model ID k3 where your membership tier supports it.
  3. 3. Test OpenRouter for API comparison: use OpenRouter's Moonshot AI listing and try moonshotai/kimi-k3 next to other models using the same prompts.
  4. 4. Save your benchmark prompts: compare coding fixes, long-context retrieval, latency, tool-call reliability, and cost before moving production traffic.

Related AI Pages

Kimi K3 FAQ

Who owns Kimi K3?+

Kimi K3 is from Moonshot AI, the company that develops the Kimi assistant and Kimi model family.

How big is Kimi K3?+

Kimi says K3 has 2.8T total parameters. That figure describes total model size, not the hardware you need to run it.

What is the Kimi K3 release date?+

Kimi Code release notes list Kimi K3 on July 16, 2026.

Is Kimi K3 open weight?+

Kimi documentation says K3 was open-sourced, and OpenRouter describes it as open-weight. Check the official model repository or license before self-hosting or commercial deployment.

How do I use Kimi K3 on OpenRouter?+

Use the OpenRouter model listing for moonshotai/kimi-k3, generate an API key, and call it through OpenRouter-compatible chat completions. A dedicated tutorial can live at /ai/tutorials/how-to-use-kimi-k3-on-openrouter.

Does Kimi K3 require a subscription?+

Kimi Code ties K3 access to membership tiers. Moderato and above can use K3, while 1M context starts at Allegretto and above in the official docs.

What hardware do I need for Kimi K3?+

There is not enough confirmed serving guidance to give a GPU count. For production planning, use the official model card, quantization notes, and serving benchmarks when available.

Is Kimi K3 better than other models?+

It may be strong for coding, agents, and long-context work, but the right answer depends on your task, latency target, price, and benchmark method.