Kimi K3 Explained: Moonshot AI's 2.8-Trillion-Parameter Open Model — Benchmarks, Pricing, and How to Access It

Kimi K3 is Moonshot AI's new flagship large language model, released on July 16, 2026. It is a 2.8-trillion-parameter open Mixture-of-Experts (MoE) model with a 1-million-token context window, priced at $3 per million input tokens and $15 per million output tokens. Moonshot says it competes with Anthropic's Claude Opus 4.8 and OpenAI's GPT-5.5 on most benchmarks, and the company has promised to release the open weights by July 27, 2026. This guide covers what Kimi K3 is, how it is built, how it performs, what it costs, how to access it, and where it sits in a remarkably crowded month of AI launches.

What is Kimi K3?

Kimi K3 is the third-generation flagship in Moonshot AI's "Kimi" family, following K2 in 2025 and the K2.5 / K2.6 point releases. It is the largest open-weight model any lab has announced to date — larger, by parameter count, than any openly released model from Meta, Mistral, DeepSeek, or Alibaba's Qwen line.

Three things make it notable:

• Scale with openness. Frontier-scale parameter counts have historically stayed behind closed APIs. Moonshot is shipping a 2.8T-parameter model and committing to publish the weights.
• Frontier-adjacent quality. On Moonshot's own evaluations, K3 mostly beats Claude Opus 4.8 and GPT-5.5 — a tier of quality Chinese open models had not previously reached.
• Aggressive pricing. At $3/$15, it is roughly the price of Claude Sonnet 5, and about one-fifth the output price of Anthropic's flagship Claude Fable 5 ($50/M output).

Kimi K3 specifications at a glance

Architecture: how Kimi K3 is built

Kimi K3 is a sparse Mixture-of-Experts model. Of its ~896 experts, only 16 are activated per token, so the active parameter count per forward pass is a small fraction of the 2.8T total — the standard MoE trade that lets a very large model run at a manageable inference cost.

Moonshot highlights two architectural ingredients:

• Kimi Delta Attention — a linearised attention variant aimed at keeping long-context inference affordable, which is what makes the 1M-token window practical rather than merely advertised.
• Attention Residuals — residual pathways around the attention blocks that Moonshot credits for stability and quality at this scale.

The practical upshot is a model designed for long-horizon, high-context work — large codebases, long documents, and multi-step agent runs — rather than short one-shot prompts.

Benchmarks: how good is Kimi K3?

On Moonshot's self-reported evaluations, K3 "mostly" beats Claude Opus 4.8 (max) and GPT-5.5 (high), while sitting below the current Western frontier — Claude Fable 5 and GPT-5.6 Sol. Independent signals so far are consistent with that framing:

• Artificial Analysis reports an Elo of ~1547 on its private evaluation set — frontier-adjacent, ahead of every previous open-weight model.
• K3 leads Arena.ai's frontend-code benchmark, an area where the Kimi family has historically been strong.
• Independent reviewer Simon Willison found K3 produced valid, spatially-coherent SVGs and handled image analysis and alt-text generation well.

Rough positioning against this month's other releases:

A caution that applies to every launch this month: vendor-reported benchmarks are directional, not gospel. Treat the numbers above as a starting hypothesis until broad independent evaluations settle.

Pricing: how much does Kimi K3 cost?

K3 is priced like a Western mid-tier model, not a flagship:

Two caveats temper the low sticker price. First, K3 is roughly 3x more expensive than K2.6 — a real jump for teams already running the older model. Second, K3 spends heavily on reasoning tokens: independent testing saw a simple task consume over 13,000 reasoning tokens, so effective cost per completed task can run higher than the headline rate suggests. Budget by task, not by token.

How to access Kimi K3

• Kimi app & web — live now for consumers.
• Playground — for quick evaluation without wiring up the API.
• API — available now; the interface follows the OpenAI-compatible conventions the Kimi family has used, so most existing SDK code needs only a base-URL and model-ID change.
• Open weights — Moonshot has promised a weights release by July 27, 2026, which will enable self-hosting and fine-tuning for teams with the hardware to run a 2.8T-parameter MoE.

Note that at launch K3 exposes a single reasoning-effort level ("max"), so unlike GPT-5.6 or Claude's effort tiers, you cannot yet dial reasoning down to save cost.

Where Kimi K3 fits in a busy month

K3 did not land in a vacuum. The four weeks before it saw the densest run of model launches in memory: Claude Sonnet 5 (June 30), Grok 4.5 (July 8), OpenAI's GPT-5.6 Sol/Terra/Luna (July 9), and Meta and Google image and translation models throughout. Against that backdrop, K3's contribution is specific: it is the first release to put frontier-adjacent quality and an open-weight promise in the same box, and to do it at a price that undercuts every Western frontier model by a wide margin. For a fuller picture, see our companion roundup of the last 30 days in AI.

Who should use Kimi K3?

• Use it now if you want frontier-adjacent quality at mid-tier prices, run long-context or agentic workloads, or need a credible open-weight option you can eventually self-host.
• Wait for the weights (July 27) if your interest is fine-tuning, on-prem deployment, or data-residency-constrained use — the hosted API is a preview of what you'll be able to run yourself.
• Stay on K2.6 if you are cost-sensitive and already happy: K2.6 is roughly a third of K3's price, and for many extraction, classification, and chat workloads the quality gap won't justify 3x spend.
• Stay on a Western frontier model (Fable 5, GPT-5.6 Sol) if you need the absolute top of the capability curve and price is secondary.

Frequently asked questions

Is Kimi K3 open source?

It is open-weight, not fully open-source. Moonshot has promised to release the model weights by July 27, 2026, which allows self-hosting and fine-tuning. That is distinct from releasing training data or a fully permissive codebase.

How big is Kimi K3?

Around 2.8 trillion total parameters in a Mixture-of-Experts design, activating 16 of roughly 896 experts per token — making it the largest open-weight model announced so far.

Is Kimi K3 better than GPT-5.6 or Claude Fable 5?

On Moonshot's own benchmarks, K3 beats Opus 4.8 and GPT-5.5 but sits below Claude Fable 5 and GPT-5.6 Sol. Independent evaluations broadly agree it is frontier-adjacent rather than outright frontier-leading.

How much does Kimi K3 cost?

$3 per million input tokens and $15 per million output tokens — about the price of Claude Sonnet 5, and roughly one-fifth the output price of Claude Fable 5.

When are the open weights released?

Moonshot has committed to publishing the weights by July 27, 2026.

The bottom line

Kimi K3 is the clearest sign yet that the open-weight frontier is closing on the closed one. It won't dethrone Claude Fable 5 or GPT-5.6 Sol at the very top, but it doesn't need to: a model that trades blows with Opus 4.8, ships a 1M-token context, costs a fraction of the Western flagships, and promises downloadable weights within two weeks is a genuinely new option for teams that care about cost, control, and openness. The one to watch is July 27 — the day the weights land and everyone can verify the benchmarks for themselves.