Kimi K3: Moonshot AI's 2.8T Open-Weight Frontier Model for Agentic Coding

Kimi K3 is Moonshot AI's new flagship large language model — and, at 2.8 trillion parameters, the largest open-weight model released to date. Launched on July 16, 2026, it pairs a frontier-class Mixture-of-Experts (MoE) design with a 1-million-token context window and an agent-first focus on coding and long-horizon knowledge work. What makes it notable isn't just the size: Moonshot is releasing the full weights, putting genuinely frontier performance in reach for self-hosted agents. Here's what K3 actually is, how it benchmarks, what it costs, and where it fits.

This guide covers K3's architecture, benchmarks, pricing, how it compares to the K2 line, and the concrete workflows builders are already running on it.

What Is Kimi K3?

Kimi K3 is an open-weight, agent-focused frontier model from Beijing-based Moonshot AI, released on July 16, 2026 as the successor to the Kimi K2 series. Moonshot describes it as the company's most powerful open-source coding model to date, built to sustain long engineering sessions, navigate large repositories, and orchestrate terminal tools with minimal human oversight. (Fortune)

The headline number is scale. K3 packs 2.8 trillion total parameters in a sparse MoE architecture — roughly 2.8× the size of K2.6, and larger than Chinese rivals like DeepSeek's V4 Pro (~1.6T) and Zhipu's GLM-5 series. That makes it, by Moonshot's account, the largest open-weight model ever built. (VentureBeat)

Crucially, the weights aren't staying locked up. Moonshot has scheduled the full open weights for July 27, 2026, meaning developers will be able to inspect, modify, and self-host the model rather than only calling a hosted API. (Axios)

Specs and Architecture

K3 keeps the OpenAI-style, agent-first design philosophy of the K2 line and pushes the ceiling on every axis.

• 2.8T-parameter sparse MoE — hundreds of experts with sparse activation, so only a fraction of the network runs per token. The scale is about breadth and specialization, tuned for long-horizon reasoning and complex tool use. (VentureBeat)
• 1-million-token context window — enough to hold an entire codebase, design system, or research corpus "in view" without elaborate chunking. (Axios)
• Kimi Delta Attention (KDA) — a hybrid linear-attention mechanism first published as open research by Moonshot, aimed at keeping ultra-long sequences tractable without the cost blow-up of standard attention. (VentureBeat)
• Attention Residuals (AttnRes) — described by Moonshot as a drop-in replacement for residual connections that delivers consistent scaling gains. (VentureBeat)
• Native multimodal input — vision understanding across text and images, plus an always-on reasoning mode Moonshot calls "thinking mode." (Axios)
• OpenAI-SDK compatible — the API mirrors the OpenAI interface, so teams already building on OpenAI or Anthropic toolchains can integrate with minimal changes. (VentureBeat)

Both KDA and AttnRes were previously published as open research on GitHub — a signal that K3's efficiency gains come from architecture, not just brute-force scale. (VentureBeat)

Benchmarks: Where K3 Lands

K3 debuted at No. 3 on the Artificial Analysis leaderboard, behind Anthropic's Claude Fable 5 and OpenAI's GPT-5.6 Sol — but ahead of everything else, and it led some practical tests outright. (Wikipedia)

A few results stand out from launch analyses:

• GDPval-AA v2 (real-world tasks across 44 occupations and 9 industries): 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). (VentureBeat)
• AA-Briefcase (a private long-horizon knowledge-work benchmark): K3 took second place at 1,527, beating GPT-5.6 Sol Max and trailing only Fable 5 Max. (VentureBeat)
• Task automation: K3 ranked first in four of eight benchmarks — including Automation Bench, SpreadsheetBench 2, and BrowseComp — finishing second to Fable 5 in most others. (VentureBeat)
• Front-end coding: in blind testing by evaluator Arena, developers preferred K3 over every leading U.S. model for front-end coding, including Fable 5 and GPT-5.6 Sol. (Axios)

One detail matters for agent builders: Moonshot says it achieved these automation results in a single-agent setup using the 1M-token context — no context compression or external context-management tricks. That hints raw context length plus strong retrieval can rival elaborate multi-agent workarounds. (VentureBeat)

The caveat: K3 has been public for only days, and viral demos and early benchmarks can overstate how reliably a model performs across real production work. Treat the leaderboard numbers as a promising starting point, not a settled verdict. (Axios)

Pricing and Access

K3 is available today via the kimi.com interface, mobile apps, and the Moonshot platform API at platform.moonshot.ai. Hosted pricing is set at roughly $3 per million input tokens and $15 per million output tokens. (MLQ)

That's the highest pricing from any Chinese AI lab — a clear break from the deep discounts Moonshot's earlier models were known for. For context, it's well above z.ai's GLM-5.2 ($4.40/M output) and DeepSeek V4 ($0.87/M output), yet still far cheaper than U.S. frontier models: Claude Fable runs around $50 per million output tokens for comparable work. (Fortune)

Two practical notes before you budget:

• K3 currently exposes only one reasoning-effort level ("max"), and independent testers report heavy reasoning-token consumption — one flagged about 13,241 tokens (roughly $0.25) to generate a single simple SVG. Long chains add up fast. (MLQ)
• Once the weights ship on July 27, self-hosting becomes an option for teams that need data residency or want to control cost at scale — the same hosted-to-self-hosted path we've seen across Moonshot's open-weight releases. (Axios)

K3 vs. K2.5 / K2.6 / K2.7: What Changed

If you've used the K2 line, K3 is less about a new API and more about raising ceilings. The platform keeps the same OpenAI-style interface and tool-calling semantics, so upgrade paths stay straightforward.

The main differentiators over K2.x: nearly 3× the parameters, a 4× longer context window (1M vs 256k), the new KDA + AttnRes architecture, and stronger multimodal reasoning. Moonshot also reports K3 uses about 21% fewer output tokens than K2.6 on equivalent tasks — continuing the efficiency trend from Kimi K2.7 Code. (MLQ)

Real-World Use Cases for Builders and Teams

K3's combination of frontier performance, huge context, and open weights makes it a candidate backbone for serious agent work:

• Autonomous coding agents that own full features — sustaining long engineering sessions, navigating massive repos, and orchestrating terminal tools with minimal supervision. (Fortune)
• Repo-scale migrations and prototyping — turning PRDs, design files, and legacy docs into working scaffolding while keeping full project context in the 1M-token window.
• Document-heavy knowledge work in finance, legal, and consulting, where a big context plus native vision helps with dense reports, and open weights let you move sensitive workflows to self-hosted infrastructure later.

This isn't hypothetical adoption. Moonshot's earlier models were already inside Western products — Cursor used Kimi to help build its Composer 2 coding agent, and DoorDash's CTO said the company delegates "lower-level work to Kimi K2.6." K3 raises the ceiling those teams are building against. (Fortune)

Why K3 Matters for the AI Landscape

Strategically, K3 is Moonshot's bid to prove open-weight models can compete head-on with closed systems in the highest-value segments: complex coding, long-horizon reasoning, and multimodal research. Analysts framed the launch as evidence that architectural innovation plus pre-training scaling can still deliver step-change gains for Chinese labs despite hardware constraints. (CNBC)

It's also a competitive shock. The release — timed just ahead of the 2026 World Artificial Intelligence Conference in Shanghai — is widely read as evidence that open source is no longer trailing closed models by many months. For builders evaluating frontier models, that means K3 belongs on the shortlist alongside Claude and GPT-5.x, especially if your roadmap includes self-hosted agents and multimodal workflows. (VentureBeat)

For the broader open-weight picture, K3 sits alongside the other frontier releases we've tracked: DeepSeek V4 Pro, Zhipu's GLM-5.2, and MiniMax-01.

Put a Model Like Kimi K3 to Work in Your Own AI Workforce

K3's real value shows up when a frontier model is wired into actual workflows — repos, documents, terminals, and multi-step plans — not just a chat box. That's exactly what a model-agnostic, multi-agent platform is for: route long-horizon coding to a model like K3, keep others for routine tasks, and orchestrate them across real work. Eigent is an open-source Cowork desktop app that runs a multi-agent AI workforce locally, so you can plug in the best model per task and keep sensitive data on your own machine. See how agents can review GitHub PRs end to end, then download Eigent to try it.

Frequently Asked Questions

What is Kimi K3?

Kimi K3 is Moonshot AI's frontier large language model, released July 16, 2026 — a 2.8-trillion-parameter sparse Mixture-of-Experts model with a 1-million-token context window and native multimodal input. Moonshot calls it the largest open-weight model ever released and its most powerful open-source coding model to date.

Is Kimi K3 open source?

Moonshot has committed to releasing K3's full weights on July 27, 2026, at which point developers can inspect, modify, and self-host it. Until then, K3 is accessible through the hosted kimi.com interface, mobile apps, and the Moonshot platform API.

How does Kimi K3 compare to Claude and GPT-5?

At launch K3 debuted at No. 3 on the Artificial Analysis leaderboard, behind Anthropic's Claude Fable 5 and OpenAI's GPT-5.6 Sol, while leading some practical tests — including Arena's blind front-end coding evaluation, where developers preferred it over both. It ranked first in four of eight task-automation benchmarks.

How much does Kimi K3 cost?

Hosted pricing is roughly $3 per million input tokens and $15 per million output tokens — the highest of any Chinese lab, but still far below U.S. frontier models like Claude Fable. K3 currently offers only a single "max" reasoning level, and testers report heavy reasoning-token usage, so long chains can get expensive.

What's new in K3 versus the K2 series?

K3 scales from K2's 1T parameters to 2.8T, extends the context window from 256k to 1M tokens, and adds two new architectural techniques — Kimi Delta Attention and Attention Residuals. It also reports about 21% fewer output tokens than K2.6 on equivalent tasks, alongside stronger multimodal reasoning.

Can I use Kimi K3 with a multi-agent platform like Eigent?

Yes. Because K3 is OpenAI-SDK compatible and (from July 27) self-hostable, a model-agnostic platform like Eigent can route long-horizon coding and knowledge-work tasks to K3 while using other models for routine steps — keeping data local when needed.