artificialwatch
model radar / moonshot / kimi-k3

Kimi K3

Moonshot AI · the open-weight heavyweight's next act
LIVE — dropped Jul 15 open weights on HF Jul 27 · 2.8T k2.6 + agent swarm already in production
Watch this model free →Alert the second it answers on a public API · no card

It happened. Kimi K3 went live July 15 — 2.8T total parameters, a 1M-token context, native vision, and a new Kimi Delta Attention + Attention Residuals architecture, at $3/M in · $15/M out. The independent verdicts came fast: Artificial Analysis scores it 57 (Opus 4.8-class, behind Fable 5 and Sol), it took #1 on Frontend Code Arena, and it went to #1 on the LMArena text board. Full open weights landed July 27 on Hugging Face — 2.8T parameters, 104B active per token, making it the largest open-weights model publicly available. Track it on the Bench Index → · the open-weights leaderboard →

kimi-k3 — settled live · jul 15
K3 released by Jul 31 — resolved YES, it landed Jul 15polymarket 100%
final pre-launch read was 97% on Jul 15 · open weights followed on Jul 27 under the bespoke Kimi K3 License — open-weight, not OSI open source
✳ the wire

Leaks & news, newest first.

Jul 27 · weights

The full 2.8T weights are on Hugging Face

Moonshot published Kimi K3’s full weights under the moonshotai organization, meeting the “by July 27” date in its launch post: 2.8T total parameters, 104B active, MXFP4 weights, under the custom Kimi K3 License — open-weight, not OSI open source. via Hugging Face · Moonshot

Jul 16 · verdict

The independent numbers: AA Index 57, #1 Frontend Code Arena — weights July 27

Artificial Analysis puts K3 at 57 on its Intelligence Index — comparable to Opus 4.8 and GPT-5.5, behind Fable 5 and GPT-5.6 Sol — with 1668 Elo on GDPval v2, #1 on AutomationBench (53%), and second only to Fable 5 on AA-Briefcase. Cost per task: $0.94 — half of Opus 4.8’s $1.80, though far above open-weights peers. Separately, K3 jumped 17 places to #1 on Frontend Code Arena (1679), first in 6 of 7 domains — second only to Fable 5 in Gaming. Moonshot confirmed the 2.8T weights ship July 27, which would make it the clear open-weights leader (GLM-5.2: 51, DeepSeek V4 Pro: 44). via Artificial Analysis · Frontend Code Arena

Jul 15 · LAUNCHED

Kimi K3 is live — 2.8T, 1M context, native vision

K3 appeared on the public API on July 15, and the official card confirms the scale: 2.8 trillion total parameters (MoE), 1M-token context, native visual understanding, and a new Kimi Delta Attention + Attention Residuals architecture, priced at $3/M input · $15/M output. Moonshot’s launch table claims a GDPval win over Opus 4.8 (max) and 91.2 on BrowseComp — lab-reported for now. Full open weights are promised within days. The chaotic launch day (leaked promo page, pulled campaign, app-picker sightings, the “Kivine” Arena mystery) resolved the way the market said it would. via Moonshot official docs · OpenRouter

Jul 15 · leak

The launch page leaked — then the launch didn't happen

A “K3 launch limited-time recharge campaign” page went live on Moonshot’s own Kimi Open Platform — promo starting Jul 15, 00:00 China time, 10–30% bonus API credits — then was pulled. The hour came and went: the official API still lists K2.7 Code and K2.6, and there’s no model card, pricing or benchmarks. A delay, an internal-campaign mixup, or a launch still landing later — nobody outside Moonshot knows. via @kimmonismus on X

Jul 15 · sighting

K3 and "K3 Agent Swarm" surface in the Kimi app picker

Screenshots circulating today show K3 selectable in the Kimi app’s model picker (“great for slides, websites, docs”) alongside a K3 Agent Swarm mode — while the API side still shows nothing. A mystery model called “Kivine” claiming Moonshot origin also appeared on Arena. Consistent with a staged consumer-first rollout; still unconfirmed by Moonshot. via community screenshots · kie.ai

late Jun · signal

Moonshot staff: "Kimi K3 is coming in July"

A Moonshot AI staff member (@Young_AGI) said K3 arrives in July — a staff post on X, not a press release, but the most direct timing signal on record. No official date from the company itself. via @Young_AGI on X

Jun · report

K3 reportedly targets multi-trillion parameters

Reports place K3 anywhere from 2.5T parameters up, aimed at reasoning, multimodal and context-length gains over K2. Ranges vary by outlet; treat the specifics as rumor until weights land. via xix.ai

Apr · baseline

K2.6 → K2.7: the open-weight bar keeps rising

K2.6 shipped in April — open weights, coding scores that tie GPT-5.5-class models, and the Agent Swarm orchestration layer now running in production. It's been topping value charts ever since. via Miraflow · Verdent

pattern · cadence

The release math points at now

Moonshot ships a major model every 2–3 months. K2.6 landed in April — the window that pattern opens is exactly the one we're in. via kimi-k2.org

✳ from moonshot

The official word.

"Today, we are introducing Kimi K3 — our most capable model."

Moonshot AI, Kimi K3 tech blog · Jul 17, 2026 — the full weights followed on Jul 27

K2.7The previous open-weight bar — K3’s 2.8T weights took that title on Jul 27.
Agent SwarmMoonshot's multi-agent orchestration layer, live today.
K3Shipped Jul 15 — 2.8T total, 104B active; open weights Jul 27.
✳ pricing

$3 / $15 per million tokens, on Moonshot AI's own API.

How much does Kimi K3 cost?

Moonshot AI's own API lists Kimi K3 (kimi-k3) at $3 per million input tokens and $15 per million output tokens, and $0.30 per million cached input tokens. Prices as published in the models.dev catalog, as of 2026-10-06.

What is Kimi K3's context window?

1.05M tokens, as Moonshot AI lists Kimi K3 (kimi-k3).

✳ where to use it

Listed by 74 providers, exact ids below.

74 providers list Kimi K3 in the models.dev catalog (as of 2026-10-06), under the model ids shown below.

Moonshot AIkimi-k3
Moonshot AI (China)kimi-k3
Amazon Bedrockglobal.moonshotai.kimi-k3 · us.moonshotai.kimi-k3
OpenRoutermoonshotai/kimi-k3
Vercel AI Gatewaymoonshotai/kimi-k3 · moonshotai/kimi-k3-fast
302.AIkimi-k3
Abacusmoonshotai/Kimi-K3
ai&moonshotai/kimi-k3
AIHubMixkimi-k3
ainetcafeKimi-K3
Alibabakimi-k3
Alibaba (China)kimi-k3

Also listed by Arcee, Baseten, Berget.AI, Charm Hyper, Chutes, ClinePass, Cloudflare AI Gateway, Cortecs, CrofAI, CrossModel, Deep Infra, DevPass (LLM Gateway), DigitalOcean, Eden AI, EmpirioLabs AI, engy, Fireworks AI, GitHub Copilot, GreenPT, Hugging Face, Impossibl, Inco, IteraCompute, Jalapeno Cloud, Kenari, Kilo Gateway, LLM Gateway, Melious, Merge Gateway, Modal, NanoGPT, Nebius Token Factory, Neon, Neuralwatt, NovitaAI, Nvidia, Ofox, Ollama Cloud, OpenCode Go, OpenCode Zen, Opper, OrcaRouter, Perplexity Agent, Pioneer, Requesty, SCNet Token Plan, SenseNova (China), SiliconFlow, Synthetic, Tempr Gateway, TensorX, Tinfoil, Together AI, TokenGo, Umans AI, Umans AI Coding Plan, Vancine, Venice AI, Vivgrid, Volcengine Ark Coding Plan, Wallaby, ZenMux.

Listings cover the model family, so a provider may list a sibling tier under the same name. Check the id before you build on it.

Head to head, on the public leaderboards:

Hear it the second K3.1 answers.

We sweep the public catalogs (OpenRouter, models.dev) every 60 seconds — and if the weights hit Hugging Face first, the custom watch catches that too. Email + Chrome push free, SMS and a phone call on paid tiers.

On the board. You'll hear the second Kimi K3.1 answers.