Kimi K3: Complete Guide to Moonshot's 2.8T Model (2026)

Kimi K3 is the newest flagship AI model from Moonshot AI, officially launched on July 16, 2026. It is the largest open-weight model ever released — a 2.8-trillion-parameter mixture-of-experts system with a 1-million-token context window, native image and video understanding, and benchmark scores that put it within striking distance of GPT-5.6 Sol and Claude Fable 5 at roughly half the cost per task.

Update (Sept 2026): K3 has since been compared with newer models: see Kimi K3 vs Claude Opus 5.5, vs GPT-6 Astra, vs Claude Fable 5.1, vs GLM-5.3-Flash and vs DeepSeek V4.1 Flash. The benchmark comparisons below are from the July launch.

This is the hub guide. Each section below gives you the essentials, then links to a dedicated deep-dive if you want more detail on that topic.

Many people search for it as “Kimi 3.0” — the official name is Kimi K3, following Kimi K2 from 2025.

What is Kimi K3?

Kimi K3 is a frontier-class large language model built by Moonshot AI, the Beijing-based lab behind the Kimi assistant. Where Kimi K2 was a strong open-weight model with a 1T-parameter architecture, K3 is a generational leap: nearly 3× the parameters, 4× the context window, and two brand-new architectural techniques — Kimi Delta Attention (KDA) and Attention Residuals — that make it dramatically faster on long contexts.

Three things define K3:

1. Scale with efficiency. 2.8 trillion total parameters, but only 16 of its 896 routed experts (plus 2 shared) are active per token, about 104B active parameters. You get frontier intelligence without frontier-level inference cost.
2. A 1M-token context window. K3 can hold entire codebases, books, or hours of video transcripts in a single prompt — and its KDA architecture makes long-context decoding up to 6.3× faster than K2.
3. Agentic strength. K3 is built for multi-step, tool-using work: coding agents, browser automation, spreadsheets, and deep research. It ranks #1 on 4 of 8 real-world agentic benchmarks in Moonshot’s launch suite.

At launch, independent testers at Artificial Analysis placed K3 at #4 out of 189 AI models (score 57), ahead of Claude Opus 4.8 and GPT-5.5 — the highest rank an open-weight model had reached at the time. Artificial Analysis has since re-versioned its Intelligence Index: on v4.3.2, K3 scores 44 and ranks #3 of 115 large (over 150B) open-weight models. The two scores are not comparable.

Key specifications at a glance

Spec	Kimi K3
Total parameters	2.8T
Architecture	KDA + Attention Residuals + Stable LatentMoE
Active parameters	~104B per token
Experts	896 routed + 2 shared (16 routed active per token)
Context window	1,048,576 tokens (1M)
Vision / video input	Native
Long-context decode	Up to 6.3× faster than K2
Scaling efficiency	~2.5× better than K2
Weights license	Custom Kimi K3 License (weights released July 27, 2026)

For the full breakdown of every spec — including what KDA and Attention Residuals actually do, and why the expert count matters — read the dedicated guide: Kimi K3 Specifications Explained.

Benchmark performance

K3’s launch benchmarks show it beating or matching the best closed models on real-world agentic tasks:

Benchmark	Kimi K3	Best competitor	Result
Program Bench	77.8	GPT-5.6 Sol (77.6)	K3 #1
SWE Marathon	42.0	Claude Opus 4.8 (40.0)	K3 #1
BrowseComp	91.2	GPT-5.6 Sol (90.4)	K3 #1
SpreadsheetBench 2	34.8	Claude Fable 5 (34.7)	K3 #1
Automation Bench	30.8	GPT-5.6 Sol (29.7)	K3 #1
Terminal-Bench 2.1	88.3	GPT-5.6 Sol (88.8)	K3 #2 by 0.5

At launch K3 also took #1 on the Frontend Code Arena leaderboard (K2.6 was #18). Moonshot itself was transparent that K3 still trailed GPT-5.6 Sol and Claude Fable 5 overall — but it was the only open model anywhere near them, and it cost far less per completed task.

For the full head-to-head — every benchmark, cost per task, and which model wins for which workload — read Kimi K3 vs GPT-5.6 vs Claude Fable 5 vs Opus 4.8.

Pricing: how much does Kimi K3 cost?

K3’s API pricing is aggressive — Sonnet-tier prices for near-frontier intelligence:

Kimi K3 API	Per 1M tokens
Cache-hit input	$0.30
Cache-miss input	$3.00
Output	$15.00
Cache write (5-min TTL, default)	$3.00
Cache write (1-hour TTL)	$6.00
Avg. cost per benchmark task (launch)	$0.94

Thanks to Moonshot’s Mooncake disaggregated inference (with >90% cache-hit rates in coding workloads), K3 works out to roughly 50% cheaper per completed task than Claude Opus 4.8 or GPT-5.5, and about 3× cheaper than Claude Fable 5 ($2.75/task).

If you just want to use K3 rather than build on it, the Kimi app has a free tier for trying Kimi, but Moonshot’s docs list K3 access only on paid plans. Since September 18, 2026 those are Go / Plus / Pro / Max (¥49 / ¥99 / ¥199 / ¥699 a month, billed monthly); K3 access starts at Plus, and legacy members keep Moderato and above.

How to access Kimi K3

There are four official Moonshot channels for K3:

1. Kimi app & kimi.com — chat on iOS, Android, HarmonyOS, or the web. The fastest way to try Kimi (K3 itself needs a paid plan).
2. Kimi Work — the desktop app (Windows / Apple silicon, v3.1.0+) for deep research, widgets and dashboards. This is the answer if you’re looking to download Kimi K3 for PC.
3. Kimi Code — terminal + IDE agent. Run it and switch to K3 with the /model command. Needs Plus or above (Pro for the full 1M context); see the Kimi Code K3 setup guide.
4. Kimi API — an OpenAI-compatible endpoint. Keys and docs are at platform.kimi.ai (the old platform.moonshot.ai redirects there); the base URL is still https://api.moonshot.ai/v1. Top up at least $1, create an API key, select kimi-k3, and you’re running.

Outside Moonshot, K3 is also in GitHub Copilot (since August 6), generally available on Amazon Bedrock (since September 18, same $3 / $15 list price), served in Microsoft Foundry via Fireworks, and offered by about 17 providers on OpenRouter. Google Vertex has no native listing. Full comparison: every Kimi K3 API provider.

from openai import OpenAI

client = OpenAI(
    api_key="YOUR_MOONSHOT_API_KEY",
    base_url="https://api.moonshot.ai/v1",
)

resp = client.chat.completions.create(
    model="kimi-k3",
    messages=[{"role": "user", "content": "Refactor this repo"}],
    reasoning_effort="max",  # low / high / max all supported
)

Two K3 model IDs in Kimi Code: the membership catalog lists k3 (up to 1M context on Pro / legacy Allegretto and above; 256K on Plus) and k3-256k (256K context). On a pay-as-you-go API key the ID is simply kimi-k3. Within 256K, k3-256k delivers the same results while consuming about half the quota — ideal for everyday Q&A, code completion, routine feature development, and single-file edits (video input is not supported on it). Both are available from Plus (legacy Moderato) up, and both support low / high / max thinking effort. Since September 11, the kimi-for-coding model in Kimi Code is served by K2.8 Preview, which Moonshot says performs close to K3.

One warning: because of the launch hype, clone and phishing sites have appeared imitating Kimi. Before entering any credentials, read Kimi AI Official Website: How to Find the Real One.

Kimi K3 vs Kimi K2

If you used K2, here’s the short version: K3 is not an update — it’s a new generation.

Spec	Kimi K2	Kimi K3
Total parameters	1.0T	2.8T
Architecture	MLA MoE	KDA + AttnRes + Stable LatentMoE
Experts	384	896 (16 active / token)
Context window	256K tokens	1,048,576 tokens
Vision / video input	—	Native
Frontend Code Arena	#18 (K2.6)	#1

The full analysis — including which model still makes sense for self-hosters and fine-tuners — is in Kimi K3 vs Kimi K2: Full Comparison.

Release timeline

• July 16, 2026 — Kimi K3 announced and live in the Kimi app, API, Kimi Work and Kimi Code.
• July 27, 2026 — open weights released on Hugging Face under the custom Kimi K3 License (details).
• August 6, 2026 — K3 generally available in GitHub Copilot.
• September 11, 2026 — K2.8 Preview takes over kimi-for-coding in Kimi Code.
• September 18, 2026 — Kimi plans relaunched as Go / Plus / Pro / Max; K3 generally available on Amazon Bedrock.

For the story of how the launch unfolded (and how the pre-launch rumors held up), see Kimi K3 Release Date & Rumor Recap.

Limitations to know about

No model is magic, and K3 has real trade-offs:

• Not #1 overall. On the Artificial Analysis Intelligence Index v4.3.2 it scores 44, behind Claude Opus 5.5 (54), GPT-6 Astra (53) and Claude Fable 5.1 (51). If you need the absolute top score regardless of cost, those still win.
• Not fast. Artificial Analysis measures about 35 tokens/s with a ~3.8s time to first token on Moonshot’s own endpoint.
• Self-hosting is heavy. The open weights are about 1.56TB and serving needs roughly 1.7TB+ of GPU memory (8×B300 or 8×MI355X-class at minimum). Most people should use the API or app. See Kimi K3 VRAM requirements.
• The free tier is limited. Moonshot’s docs list K3 access only on paid Kimi plans (Plus and above), so the free tier is for trying Kimi, not a free K3.
• Reasoning effort defaults to max in the API — great for quality, but drop to low/high (or, in Kimi Code, switch to the k3-256k variant at ~half quota) for simple tasks.

FAQ

Is Kimi K3 free? Partly. The Kimi app and kimi.com have a free tier for trying Kimi, but Moonshot’s docs list K3 access only on paid plans: Plus and above since September 18, 2026, or legacy Moderato and above (the full 1M context in Kimi Code needs Pro). Moonshot doesn’t say which model free users get. The API is pay-as-you-go and unlocks after a $1 top-up.

Is Kimi K3 open source? It is open-weight. The weights were released on July 27, 2026 on Hugging Face under the custom Kimi K3 License (not the modified MIT license used for K2). Commercial use, fine-tuning and redistribution are allowed, but Model-as-a-Service businesses with more than $20M revenue over any 12 months need a separate agreement with Moonshot, and very large products must display the Kimi K3 name. Check the license before commercial deployment.

What’s the difference between “Kimi K3” and “Kimi 3.0”? Nothing — they’re the same model. “Kimi K3” is the official name; “Kimi 3.0” is just how many people search for it.

Is Kimi K3 better than GPT-5? On several real-world agentic benchmarks at launch (Program Bench, SWE Marathon, BrowseComp, SpreadsheetBench 2, Automation Bench), yes — K3 scored #1. On aggregate rankings it trails the top closed models: on the Artificial Analysis Intelligence Index v4.3.2 it scores 44, behind Claude Opus 5.5 (54), GPT-6 Astra (53) and Claude Fable 5.1 (51), slightly ahead of GPT-6 Sol (43). At launch it cost roughly half as much per task as Claude Opus 4.8 or GPT-5.5.

The bottom line

Kimi K3 is the largest open-weight model released so far and still one of the most capable (#3 among large open-weight models on Artificial Analysis’ current index). At launch it was the first open model that genuinely pressured the closed frontier on real-world agentic work — at half the price. Whether you’re a developer looking at the API, a researcher working with the open weights released July 27, or a curious user who just wants to try it in the app, the guides linked above will take you deeper on every topic.