Kimi K3 is Moonshot AI's flagship model and the first open-weight model in the 3-trillion-parameter class: 2.8 trillion total parameters with 104 billion active per token, a 1M tokens context window, native vision and video, and pricing of $3.00 per million input tokens and $15.00 per million output — the same rate as Claude Sonnet 5, for a model that competes with tiers priced well above it.
It reached the API on 16 July 2026 and the weights were published on 27 July 2026. This is the hub for everything we have written about it: what it costs, what it is good at, how it compares with both its predecessor and the closed frontier models, and the configuration details that catch people out on day one.
Specifications
| Spec | Kimi K3 |
|---|---|
| Model ID (API) | kimi-k3 |
| Developer | Moonshot AI |
| Parameters | 2.8 trillion total, 104 billion activated per token |
| Experts | 896 total, 16 selected per token (Stable LatentMoE) |
| Attention | Kimi Delta Attention (KDA) with Gated MLA, across 93 (69 KDA + 24 Gated MLA + 1 dense) layers |
| Context window | 1,048,576 tokens (1M) |
| Max output | 131,072 tokens by default, configurable to 1,048,576 tokens |
| Price | $3.00 input / $15.00 output per million tokens |
| Cached input | $0.30 per million tokens |
| Reasoning | Always on; reasoning_effort low, high, max (default max) |
| Vision | Native — MoonViT-V2 (401M), images and video |
| Open weights | Yes — MXFP4 / MXFP8, ~1.4 TB, Kimi K3 License |
| API compatibility | OpenAI-compatible and Anthropic-compatible |
What Kimi K3 is actually for
K3 is aimed at complex coding, long-context knowledge work, and long-horizon agentic workflows — jobs where a model has to stay coherent across many steps and finish what it started. It is deliberately not a general-purpose default for everything: because reasoning cannot be switched off, high-volume trivial work is genuinely better served by a smaller model.
Four characteristics define working with it day to day:
- It always reasons. There is no off switch, which sets a quality floor on hard tasks and a cost floor on easy ones.
- The default effort level is the expensive one.
reasoning_effortdefaults tomax. Setting it explicitly is the first thing to do. - Caching is where the economics live. $0.30 against $3.00 per million input tokens is a 10x discount, applied automatically to a stable prompt prefix.
- The weights are real. ~1.4 TB of them, which is both a genuine option and a much larger commitment than "open weights" usually implies.
The headline result
K3's most cited benchmark is a coding one: first place on the Frontend Code Arena at 1,679 points in blind developer voting, ahead of Claude Fable 5 — a model priced at $10 / $50 against K3's $3.00 / $15.00. On third-party aggregate rankings it places in the top five overall, behind Claude Fable 5 and GPT-5.6 Sol. Our benchmarks page has every published score with sourcing and the caveats that go with them.
Use cases
| Guide | What it covers |
|---|---|
| Kimi K3 for Coding: Setup, Settings and Real Limits | Working with K3 for agentic coding |
| Kimi K3 for AI Agents: Long-Horizon Runs and Swarms | Working with K3 for long-horizon agents |
| Kimi K3 for Long Documents: Using the Full 1M-Token Window | Working with K3 for long-document analysis |
| Kimi K3 for Deep Research: Search-Heavy Work at Scale | Working with K3 for deep research |
| Kimi K3 for Frontend Development: The Arena Winner in Practice | Working with K3 for frontend and UI work |
| Kimi K3 for Vision: Images, Video and Charts | Working with K3 for image and video understanding |
| Kimi K3 for Self-Hosting: What It Actually Takes | Working with K3 for self-hosted deployment |
| Kimi K3 for Data Analysis: Spreadsheets, SQL and Charts | Working with K3 for data analysis |
| Kimi K3 for Code Review: Catching What CI Misses | Working with K3 for code review |
Comparisons
| Comparison | The question it answers |
|---|---|
| Kimi K3 vs Kimi K2: What Actually Changed | Kimi K3 against Kimi K2 |
| Kimi K3 vs Claude Sonnet 5: Identical Price, Different Bets | Kimi K3 against Claude Sonnet 5 |
| Kimi K3 vs Claude Opus 5: Open Weights at 60% of the Price | Kimi K3 against Claude Opus 5 |
| Kimi K3 vs Claude Fable 5: The Arena Upset, In Context | Kimi K3 against Claude Fable 5 |
| Kimi K3 vs GPT-5.6 Sol: Half the Output Price | Kimi K3 against GPT-5.6 Sol |
| Kimi K3 vs ChatGPT: A Model Against a Product | Kimi K3 against ChatGPT |
| Kimi K3 vs Gemini: Open Weights vs Google Integration | Kimi K3 against Gemini |
Specs, pricing and setup
| Guide | What it covers |
|---|---|
| Kimi K3 Pricing Explained (2026) | Kimi K3 costs $3.00 per million input tokens and $15.00 per million output, with cached input at $0.30. What that means in practice. |
| Kimi K3 Context Window: 1M Tokens, Explained | What Kimi K3's 1,048,576-token context window holds in pages, code, and cost — and how to use it without wasting money. |
| How to Access Kimi K3: Every Route in 2026 | Apps, API, third-party routers, and self-hosting — every way to use Kimi K3, and which one fits you. |
| Kimi K3 API: A Practical Getting-Started Guide | Base URL, model ID, parameters, and the four API details that catch people out when they first call Kimi K3. |
| Kimi K3 reasoning_effort: low, high, max — Which to Use | Kimi K3's reasoning_effort defaults to max, the most expensive setting. What the three levels change and how to choose. |
| Kimi K3 Benchmarks: Every Published Score | Kimi K3's published benchmark results — GPQA Diamond 93.5, Terminal-Bench 88.3, Frontend Code Arena 1st — and how to read them. |
| Kimi K3 Architecture: 2.8T Parameters, 104B Active | How Kimi K3 works: Stable LatentMoE with 896 experts, Kimi Delta Attention, MXFP4 quantisation-aware training. |
| Kimi K3 Open Weights and Licence: Read This First | Kimi K3's weights are downloadable under a custom licence — not MIT. The revenue thresholds and attribution rules that apply. |
| Kimi K3 Hardware Requirements: An Honest Assessment | Can you run Kimi K3 locally? At ~1.4 TB of weights, almost certainly not. What self-hosting a 2.8T model actually requires. |
| K3 Max vs K3 Swarm Max: Which Variant Do You Need? | Kimi K3 shipped in two serving variants. K3 Max handles chat and single agents; K3 Swarm Max handles parallel multi-agent work. |
| Kimi K3 Tool Calling: Function Calls and Dynamic Loading | How function calling works on Kimi K3, including dynamic tool loading — and why offering fewer tools produces better results. |
| Is Kimi K3 Free? What You Actually Pay | Kimi K3 has free consumer app access and paid API access at $3/$15 per million tokens. The weights are free to download but not free to run. |
| Kimi K3 Prompt Caching: The 10x Discount Explained | Cached input on Kimi K3 costs $0.30 per million tokens against $3.00 standard. How to structure prompts to actually get it. |
| How to Cut Kimi K3 Costs Without Losing Quality | Four levers that reduce Kimi K3 spend — caching, effort levels, output discipline, and routing — ranked by impact. |
| Migrating to Kimi K3 from OpenAI or Anthropic | Kimi K3 speaks both OpenAI and Anthropic API shapes, so migration is mostly a base URL change. What actually needs re-tuning. |
| Kimi K3 Max Output: 131K by Default, 1M if You Ask | Kimi K3's max_completion_tokens defaults to 131,072 and can be set as high as 1,048,576 — the full context window as output. |
| Kimi K3 Limitations: What It Can't Do | An honest list of Kimi K3's constraints — no reasoning off switch, no image generation, restrictive licence, and more. |
| Kimi K3 Alternatives: 5 Models Worth Comparing | The realistic alternatives to Kimi K3 — Claude Sonnet 5, Claude Opus 5, GPT-5.6 Sol, Claude Fable 5, and Kimi K2 — and when each wins. |
Where Kimi K3 sits on price
| Model | Price (per 1M tokens) | Use it for |
|---|---|---|
| Claude Fable 5 | $10 / $50 | The hardest long-horizon autonomous work |
| GPT-5.6 Sol | $5 / $30 | Ingestion beyond 1M tokens; Azure-first stacks |
| Claude Opus 5 | $5 / $25 | Complex coding and deep reasoning on a mature enterprise stack |
| Kimi K3 | $3.00 / $15.00 | Frontier-scale capability at mid-tier pricing, with open weights |
| Claude Sonnet 5 | $3 / $15 | The same price, with reasoning you can disable and multi-cloud availability |
| Claude Haiku 4.5 | $1 / $5 | High-volume, well-specified, latency-sensitive tasks |
Most teams that get this right do not pick one model. They route: a cheap model by default, K3 on escalation, and a measured evaluation set deciding where the boundary sits. Competitor figures above are consistent with those used across our Claude Opus 5 coverage.
How to get it
Four routes — the Kimi apps, the Kimi API at https://api.moonshot.ai/v1, third-party routers, and self-hosting the open weights. Because the API is both OpenAI- and Anthropic-compatible, trying it is usually a base URL change rather than a migration. Our access guide covers which route fits you, and the API guide covers the four defaults worth changing before you ship.
Ready to go deeper?
Compare AI chatbotsFrequently Asked Questions
What is Kimi K3?
Kimi K3 is Moonshot AI's flagship model and the first open-weight model in the 3-trillion-parameter class: 2.8T total parameters with 104B active per token, a 1,048,576-token context window, native vision and video, and always-on reasoning. It reached the API on 16 July 2026 and the weights were published on 27 July 2026.
How much does Kimi K3 cost?
$3.00 per million input tokens and $15.00 per million output — the same rate as Claude Sonnet 5, 40% below Claude Opus 5, and about a third of Claude Fable 5. Cached input drops to $0.30 per million, a 10x discount applied automatically to a stable prompt prefix.
Is Kimi K3 open source?
It has open weights, which is not quite the same thing. They are published under a custom Kimi K3 License rather than MIT: commercial use is broadly permitted, but Model-as-a-Service operators above $20M in revenue over any consecutive 12 months need a separate agreement, and products above 100 million monthly active users must display "Kimi K3" in the interface.
Can I run Kimi K3 on my own hardware?
Only with a multi-node GPU cluster. The weights are roughly 1.4 TB and all 2.8 trillion parameters must be resident in memory to serve a token — MoE sparsity reduces compute per token, not memory. For almost everyone the API is dramatically cheaper than self-hosting.
What is Kimi K3 best at?
Complex coding — it placed first on the Frontend Code Arena at 1,679 points in blind developer voting, ahead of Claude Fable 5 — plus long-horizon agentic work, long-context document analysis, and search-heavy research. It is a poor fit for high-volume trivial work, because reasoning cannot be disabled.
