Kimi K3
Moonshot AI's 2.8T-parameter multimodal chatbot with a 1M-token context window, tuned for coding and long-horizon agentic tasks.
Updated 2026-07-17
Overview
Kimi K3 is Moonshot AI's flagship chat assistant, powered by a 2.8-trillion-parameter (mixture-of-experts) multimodal model with a 1-million-token context window. It's positioned squarely at the frontier tier — the July 16, 2026 launch put it near the top of several frontend-coding leaderboards, where Moonshot pitches it directly against Claude Fable 5 and GPT-5.6 Sol. You reach it through the kimi.com web app and mobile clients, or via an OpenAI-compatible API for developers.
The pitch is agentic work over very long inputs: reasoning across whole codebases, multi-file refactors, and multi-step tool use without losing the thread. The 1M context is the headline, but the more interesting story is price — Moonshot has consistently undercut US frontier labs, and K3's API rate of $3 in / $15 out per million tokens lands well below comparable Claude and GPT tiers while claiming similar coding scores. That combination is what's driving the attention, especially among developers who found earlier Kimi releases capable but rough around the edges in English.
Who's it for? Developers and power users who want a long-context, coding-strong assistant and care about token economics — and who are comfortable with a Chinese lab's data-handling and availability tradeoffs. Benchmark leadership at launch is common and rarely durable, so treat the leaderboard claims as a starting point rather than a settled verdict.
Key features
1M-token context
Holds roughly a million tokens in a single session, enough to load large codebases or long document sets without chunking, which matters for whole-repo reasoning and extended agent runs.
Agentic coding
Tuned for multi-step tool use and multi-file code changes; Moonshot's launch benchmarks emphasized frontend coding tasks where K3 topped several leaderboards at release.
Multimodal input
Accepts images alongside text so you can feed screenshots, diagrams, or UI mockups into the same conversation as code and prose.
Low-cost API
OpenAI-compatible endpoint priced at $3 input / $15 output per million tokens, undercutting comparable frontier tiers from US labs for high-volume agent workloads.
Pricing
Free tier: Free access to Kimi K3 with limits on messages and long-context usage.
| Plan | Price | What's included |
|---|---|---|
| Free | $0 | Access to Kimi K3 via web and mobile with usage limits on messages and long-context requests. |
| Paid | From $19/mo | Higher rate limits, priority access, and expanded long-context usage. Check website for exact tier breakdown. |
| API | $3 / $15 per M tokens | OpenAI-compatible API billed per token (input / output) for developers building on K3. |
Access to Kimi K3 via web and mobile with usage limits on messages and long-context requests.
Higher rate limits, priority access, and expanded long-context usage. Check website for exact tier breakdown.
OpenAI-compatible API billed per token (input / output) for developers building on K3.
Pros & cons
Pros
- ✓1M-token context window handles whole codebases and long document sets in one session
- ✓Strong frontend-coding benchmark results at launch, competitive with Claude Fable 5 and GPT-5.6 Sol
- ✓API pricing ($3/$15 per M tokens) undercuts comparable US frontier tiers
- ✓Multimodal input plus an OpenAI-compatible API make it easy to drop into existing tooling
- ✓Generous free tier for trying it before committing
Cons
- ×Launch-day benchmark leadership rarely holds as rivals ship and independent testing catches up
- ×Chinese-lab data handling, content policies, and regional availability may be dealbreakers for some teams
- ×Exact paid-tier limits and long-context caps aren't fully transparent — check the site
- ×As a brand-new release, real-world reliability under sustained agentic workloads is still unproven
How it compares
| Tool | Best for | Pricing | Score |
|---|---|---|---|
| Kimi K3 | — | Free tier + paid from $19/mo; API $3/$15 per M tokens | 8.5/10 |
| ChatGPT vs ChatGPT → | — | Free tier + Plus $20/mo + Pro $200/mo | 9.5/10 |
| Claude vs Claude → | — | Free tier + Pro $20/mo + Team $30/mo/user | 9.5/10 |
| Gemini vs Gemini → | — | Free tier + Advanced $19.99/mo | 9.2/10 |
Compare head-to-head
Related reading
Kimi K3: Moonshot's 2.8T Open Model Launches
Moonshot AI launched Kimi K3, a 2.8-trillion-parameter open-weights model with 1M context. API is live; weights drop July 27.
OpenAI's GPT-Red: Self-Improving AI Safety Explained
OpenAI's GPT-Red trains an AI to attack other models with self-play, and OpenAI says it cut GPT-5.6 Sol prompt-injection failures 6x.
Thinking Machines Inkling: Murati's First Open Model
Thinking Machines shipped Inkling, Mira Murati's first open-weight model: a 975B multimodal release built for fine-tuning, not benchmark wins.
Ready to try Kimi K3?
Head to the official site to start with Kimi K3 — pricing and plans are listed above.
Visit Kimi K3