Book a Free Strategy Call
Skip the read: talk to Walid in 30 min.
Free strategy call. We map your AI engineering team, you keep the notes.
What Is Kimi K3? Moonshot AI's New Flagship Model Explained
Kimi K3 is Moonshot AI's new flagship large language model, released the week of July 14, 2026. It replaces K2.5 as the company's top-tier model, and Moonshot is positioning it around long-horizon coding, knowledge work, and deep reasoning rather than chat.
This guide covers what K3 actually is: the specs Moonshot has published, how it differs from the rest of the K-series lineup, how to access it, and what's still unknown, including pricing and independent benchmarks.
TL;DR
- Kimi K3 launched the week of July 14, 2026, per Moonshot AI's own announcement. Coverage differs by a few days on the exact date, so treat "mid-July 2026" as the safe window.
- 2.8 trillion parameters, a 1 million token context window, and a maximum output of 131,072 tokens.
- Natively multimodal: text, images, and video.
- Uses MXFP4 quantization.
- K3 is not open-weight. Moonshot's latest open release remains K2.5; there is no K3 repository on the company's GitHub.
- Access is via API at
api.moonshot.cn/v1, in an OpenAI-compatible request format. - Pricing is not fully public. Moonshot lists separate input and output token rates but hasn't published the exact per-token numbers. File operations are "temporarily free."
- No official K3 benchmark suite has been published yet. Don't trust any benchmark number you see attributed to K3 right now, including on this site.
Related Reads
What Is Kimi K3?
Kimi K3 is the new flagship model in Moonshot AI's Kimi lineup, sitting above K2.7 Code and K2.6. Moonshot positions it for long-horizon coding, knowledge work, and deep reasoning: tasks that require sustained context and multi-step follow-through rather than single-turn chat replies.
The model is large: Moonshot and Hugging Face both list K3 at 2.8 trillion parameters. It ships with a 1 million token context window and a maximum output of 131,072 tokens per response, according to Moonshot's own API documentation. K3 is also natively multimodal, handling text, images, and video inputs directly rather than through a separate vision adapter, per Moonshot's overview docs.
On the technical side, K3 uses MXFP4 quantization, a detail Moonshot and Hugging Face published alongside the release.
Free weekly brief
Steal our production automations
The exact n8n flows, Claude Code setups, and prompts we ship for clients, broken down step by step. No spam, unsubscribe anytime.
The K-Series Lineup: Where K3 Fits
K3 is the flagship, but it isn't Moonshot's only current model. As of this writing, the active lineup is:
| Model | Role | Context window |
|---|---|---|
| Kimi K3 | Flagship | 1M tokens |
| Kimi K2.7 Code | Code-focused | 256K tokens |
| Kimi K2.6 | General | 256K tokens |
| Kimi K2.5 | Latest open-weight release | (prior generation) |
K2.5 matters for a specific reason: it's the newest model Moonshot has actually open-sourced. K3 is not on github.com/moonshotai, and there's no indication yet that it will be. If open weights matter for your use case, K2.5 (not K3) is the model to look at today.
Is Kimi K3 Open-Weight?
No. K3 is closed. Moonshot has not published K3's weights, and there's no K3 repository on the company's GitHub. This is a real shift from the framing some outlets have used around Moonshot's China-based, open-source-friendly reputation. K2.5 remains the company's most recent open release, and K3 sits apart from it as an API-only flagship.
How to Access Kimi K3
K3 is available through Moonshot's API. The base URL is api.moonshot.cn/v1, and the request format is OpenAI-compatible, so if you already have code that talks to an OpenAI-style chat completions endpoint, adapting it is mostly a matter of changing the base URL, API key, and model name rather than rewriting your integration.
from openai import OpenAI
client = OpenAI(
base_url="https://api.moonshot.cn/v1", # Moonshot's OpenAI-compatible endpoint
api_key="YOUR_MOONSHOT_API_KEY",
)
response = client.chat.completions.create(
model="kimi-k3", # confirm the exact model string in your Moonshot console
messages=[
{"role": "user", "content": "Summarize this document and flag any open questions."},
],
)
print(response.choices[0].message.content)
Because the interface mirrors OpenAI's, most existing SDKs and tooling built for OpenAI-compatible APIs should work with only configuration changes. Confirm the exact model identifier string in your own Moonshot console before shipping, since Moonshot hasn't published a single canonical model name in its public docs at the time of writing.
Kimi K3 Pricing: What's Actually Known
Here's where we have to be honest about the gap. Moonshot's pricing page lists separate input and output token pricing for K3, and notes that file operations are "temporarily free." What it doesn't list is the exact per-token rate.
We are not going to invent a number. Until Moonshot publishes exact rates, treat K3 pricing as unknown. One data point worth flagging: Pingwest, a China-focused tech outlet, characterized K3 as "very strong, very expensive." That's Pingwest's own characterization, not a confirmed price point, and we're citing it as their view rather than our claim.
If you need firm numbers to budget against, wait for Moonshot's official rate card rather than estimating from early coverage.
Are There Kimi K3 Benchmarks?
No official benchmark suite for K3 has been published as of this writing. That matters because K3 is entering a news cycle full of comparison claims ("kimi k3 vs claude fable 5," "kimi k3 vs gpt-5.6"), and none of them can currently point to a verified, apples-to-apples number for K3.
Anywhere you see a specific K3 benchmark score right now, including comparisons against Claude Fable 5, GPT-5.6, or anything else, treat it as unverified until Moonshot or an independent evaluator publishes one. We're not citing any K3 benchmark number in this post for that reason. See our Kimi K3 vs Claude Fable 5 comparison for how we handle that gap when the comparison question comes up directly.
Why Kimi K3 Is Getting Attention
Coverage of K3 has clustered around a few angles. Forbes framed the launch as a "will it challenge OpenAI and Anthropic" story. TechCrunch's framing has been more skeptical in tone ("threat or menace"). There's also a Chinese AI independence narrative running through coverage, and Yahoo Finance reported the launch putting pressure on AMD's stock. We're not able to independently verify the stock-pressure claim's mechanics here; we're noting it as a reported media angle, not a financial analysis of our own.
Bottom Line
Kimi K3 is a real, large, natively multimodal flagship model from Moonshot AI, with a 1 million token context window and specs that put it in serious flagship territory on paper. What it isn't, yet, is fully priced or independently benchmarked. The honest state of play in mid-July 2026: strong specs, a closed-weights model (K2.5 stays the open option), an OpenAI-compatible API you can start testing today, and two open questions, price and performance, that neither Moonshot nor anyone else has fully answered in public yet.
If you're evaluating K3 against Claude's current flagship, see our Kimi K3 vs Claude Fable 5 breakdown, and for the full picture of Anthropic's lineup, our model comparison hub.
FAQ
When was Kimi K3 released? The week of July 14, 2026. Moonshot AI's own homepage points to July 14, while China Money Network's coverage cites July 17. Given the discrepancy, "mid-July 2026" or "the week of July 14" is the accurate way to describe the timing without picking a single exact day.
How many parameters does Kimi K3 have? 2.8 trillion, per Moonshot AI and a Hugging Face blog post accompanying the release.
What is Kimi K3's context window? 1 million tokens for input context, with a maximum output of 131,072 tokens per response, according to Moonshot's API documentation.
Is Kimi K3 open-source? No. K3 is not open-weight and is not published on Moonshot's GitHub. Kimi K2.5 remains Moonshot's most recent open-weight release.
How much does Kimi K3 cost? Exact per-token pricing has not been published. Moonshot's pricing page confirms separate input and output rates exist and that file operations are temporarily free, but doesn't list the numbers. Treat any specific price you see elsewhere as unconfirmed until Moonshot publishes a rate card.
Are there Kimi K3 benchmarks? No official benchmark suite has been published yet. Any K3 benchmark number circulating right now, including in comparison posts, should be treated as unverified.
How do I access Kimi K3?
Through Moonshot's API at api.moonshot.cn/v1, using an OpenAI-compatible request format. You'll need a Moonshot API key from their developer console.
Sources
- Moonshot AI homepage (moonshot.ai), K3 launch announcement
- China Money Network, K3 coverage (July 17, 2026)
- Moonshot AI + Hugging Face blog, K3 model card (parameters, quantization)
- platform.kimi.com/docs/api/chat (context window, max output)
- platform.kimi.com/docs/overview (multimodal capabilities)
- platform.kimi.com/pricing (pricing structure)
- github.com/moonshotai (open-weight release history)
- Pingwest, K3 pricing characterization
- Forbes, TechCrunch, Yahoo Finance (media coverage angles)
Continue Reading
Kimi K3 Open Weights Are Live: The License Catch
Moonshot published Kimi K3's full open weights on July 27, 2026, first open model over 3T parameters, under a license with a $20M Model-as-a-Service revenue clause. Here is what it means for self-hosting vs. Claude.
Claude Cowork vs Claude Code: Which One Your Team Actually Needs (2026)
Claude Cowork and Claude Code run on the same models but solve different jobs. A decision framework, full feature and pricing comparison, and where an implementation partner fits.
How to Set Up Claude Cowork the Right Way (Setup Guide + Safety Notes)
A step-by-step Claude Cowork setup guide: prerequisites, the right first task, common mistakes, and the permissions guardrails worth understanding before you connect anything irreplaceable.
Book a Free Strategy Call
Building this in production?
Walid runs a 30-min call to map your AI engineering team. Free, no slides.
Free weekly brief
Steal our production automations
The exact n8n flows, Claude Code setups, and prompts we ship for clients, broken down step by step. No spam, unsubscribe anytime.

Ex-IBM AI engineer and enterprise architect. Adel owns the technical architecture behind every automation and AI agent system AY Automate ships.



