Kimi K3: Moonshot's 2.8T Multimodal Model Explained

On July 16, 2026, Moonshot AI released Kimi K3 — a 2.8 trillion parameter model the Beijing lab calls the first open 3T-class system, and the largest open-weight model anyone has ever shipped. Within a day, it was topping a frontend coding leaderboard ahead of Claude Fable 5 and rattling stock tickers in a way not seen since the DeepSeek moment of early 2025. Kimi K3 is not a curiosity. It is a premium-tier, multimodal reasoning model that is already live on CoreAI's model library, and if you write code or run agents, you should know what it can do.

What Is Kimi K3?

Kimi K3 is Moonshot AI's new flagship: a mixture-of-experts reasoning model with 2.8 trillion total parameters, native vision, and a context window of 1,048,576 tokens. That context size puts it in the small club of models that can hold an entire codebase, a legal archive, or a season of meeting transcripts in a single conversation — company it shares with Claude Sonnet 5 and GPT-5.6, as we covered in our guide to 1M-token context models.

The headline word is "open-weight." Moonshot says full weights are due by July 27, which means researchers and companies will be able to run the largest model ever released on their own hardware — assuming they have a small data center lying around. For everyone else, the hosted version is the practical route, and it is where the pricing story gets interesting.

The Architecture: 896 Experts, 16 Awake at a Time

A 2.8T dense model would be absurd to serve. Kimi K3 avoids that with an aggressive mixture-of-experts design: 896 experts, of which just 16 activate per token. That is about 1.8% of the pool doing the work at any moment, which is how a 3T-class model ends up with per-token prices closer to a mid-size flagship than a monster.

Moonshot also introduced two architectural changes it credits for the scaling gains: Kimi Delta Attention, a hybrid linear attention mechanism, and Attention Residuals, a drop-in replacement for standard residual connections. The nerdy details are in the launch coverage at Tom's Hardware; the practical upshot is a model that stays coherent deep into very long contexts — exactly what you want for agentic work that runs for hundreds of steps.

How Does Kimi K3 Score on Benchmarks?

Moonshot was unusually honest at launch: the company said K3 still sits behind Claude Fable 5 and GPT-5.6 Sol on overall performance. But under that ceiling, the results are striking. In Moonshot's own evaluation suite, K3 beat every other model — including Claude Opus 4.8 and GPT 5.5 — across coding and agentic benchmarks.

Independent numbers back that up. Arena ranked K3 first in its Frontend Code evaluation at 1,679 points, ahead of Claude Fable 5, in blind developer testing. On GDPval-AA v2, a benchmark for real-world knowledge work, Kimi K3 scored 1,687 — third overall behind Claude Fable 5 Max (1,815) and GPT-5.6 Sol Max (1,747.8), and ahead of Claude Opus 4.8 (1,600). For an open-weight model, sitting third on a knowledge-work leaderboard dominated by closed premium flagships is the story. Simon Willison's hands-on notes are worth a read for a practitioner's view.

Kimi K3 Pricing and How It Compares

Per-token API pricing lands at $3 per million input tokens and $15 per million output tokens, with cache-hit input at $0.30 per million. That is premium-tier money — but look at what it undercuts.

Against closed premium flagships like Claude Fable 5 or GPT-5.6 Sol, K3 is meaningfully cheaper while landing within reach on the boards that matter for coding. Against standard-tier models it is a deliberate splurge — you pay up for the 2.8T ceiling. Our honest take: if your work is agentic coding, frontend generation, or long-horizon automation, K3 is worth the premium slot. For everyday chat and writing, a standard-tier model will not feel slower or dumber, and it costs a fraction as much.

Should You Use Kimi K3 for Coding and Agents?

Yes — that is precisely where it earns its keep. Moonshot built K3 for complex coding, knowledge work, and long-horizon agentic workflows, and the Frontend Code result suggests the training worked. The million-token window means an agent can carry an entire repository plus its own scratch notes without forgetting step 12 by step 200. The vision input means you can paste a screenshot of a broken layout and ask it to fix the CSS.

It also extends a pattern we flagged months ago: Chinese labs are no longer trailing on the workloads enterprises actually pay for. We dug into that shift in our piece on Chinese AI models winning enterprise share, and K3 is the strongest data point yet.

How to Try Kimi K3 Without an API Key

You do not need a developer account or per-token billing to use it. Kimi K3 is live in CoreAI alongside 300+ other models — same chat interface on iOS, Android, and the web, one subscription, with plans from $57/month. The genuinely useful move is to open Compare and run your own prompt through Kimi K3 and Claude Fable 5 side by side. Benchmarks are averages; your workload is not. Two minutes of side-by-side output on your actual task will tell you more than any leaderboard, and switching models later is a dropdown, not a migration.

Frequently Asked Questions

Is Kimi K3 really the largest open-weight model ever?

Yes. At 2.8 trillion total parameters, Kimi K3 is the largest open-weight release to date, and Moonshot describes it as the first open 3T-class system. Full weights are due by July 27, 2026. Because only 16 of 896 experts activate per token, serving costs are far below what the parameter count implies.

What can Kimi K3 accept as input?

Text and images, with text output. Native vision means it can read screenshots, diagrams, and UI mockups directly — useful for frontend debugging and document-heavy work. Combined with the 1,048,576-token context window, it can process very large mixed inputs in one pass.

How much does Kimi K3 cost?

Via API, $3 per million input tokens and $15 per million output tokens, with cached input at $0.30 per million. On CoreAI it is included as a premium-tier model under one subscription, so you can use it without an API key or per-token billing.

Is Kimi K3 better than Claude Fable 5?

On overall performance, no — Moonshot itself says K3 trails Claude Fable 5 and GPT-5.6 Sol. But it beat Fable 5 in Arena's blind Frontend Code evaluation (1,679 points) and outscores Claude Opus 4.8 on the GDPval-AA v2 knowledge-work benchmark. For coding-heavy and agentic tasks, it is a genuine rival at a lower price.

Where can I try Kimi K3 right now?

It is live on CoreAI's model library across iOS, Android, and the web. You can chat with it directly or benchmark it against other flagships with your own prompts in Compare mode.