New

Seedream 5.0 Pro is here — ByteDance's most powerful image editing model, now on Fullmira.

Try it free
Released July 16, 2026 · 2.8T parameters · Open weights

Use Kimi K3 free — the largest open-weight model

Moonshot AI's 2.8-trillion-parameter flagship, ranked #1 on Arena's Frontend Code leaderboard. 1M context, native vision, open weights. Try it here, then open the full studio.

Kimi K3 · Max
I'm Kimi K3, running at max reasoning effort — currently the only level Moonshot exposes. I'm strongest on frontend code, long documents and agentic work. Fair warning: I think hard before answering, so give me something worth the effort.
Typically 1–5 credits per message, based on length
Want file uploads, saved history and every other model?Open Kimi K3 in the full studio
No watermarkNo credit cardSwitch models anytime
At a glance

The first open model that competes on capability

Moonshot shipped K3 on July 16, 2026. It's a mixture-of-experts design that routes each token through just 16 of 896 experts — roughly 1.8% of the pool.

2.8TTotal parameters
16 / 896Experts active per token
1MContext window
#1Frontend Code Arena
Model comparison

Kimi K3 vs Claude Fable 5 and GPT-5.6 Sol

Independent and self-reported figures side by side, including the places K3 comes second.

BenchmarkKimi K3Claude Fable 5GPT-5.6 SolClaude Opus 4.8
Frontend Code Arena· blind developer votes1,6792nd
GDPval-AA v2· 44 occupations, real tasks1,6871,8151,747.8
Artificial Analysis Elo· long-horizon knowledge work1,5471st
SWE Marathon· long-running engineering1st
Program Bench· program synthesis1st
Terminal-Bench 2.1· command-line agents2nd (−0.5)1st
Relative cost / task· Artificial Analysis est.1.0×1.1×1.9×
Context window· input tokens1M

Figures from Moonshot's K3 announcement, Artificial Analysis, Arena.ai and Simon Willison's write-up. The honest picture: Moonshot's own benchmarks show K3 mostly losing to Claude Fable 5 and GPT-5.6 Sol overall. The cost row is a relative estimate normalised to K3.

Capabilities

What Kimi K3 is actually good at

2.8T parameters, sparsely used

896 expert subnetworks with only 16 active per token. Enormous capacity without paying full-dense compute on every request — the design that makes a model this size practical to serve.

#1 on Frontend Code Arena

In blind developer voting it ranked first at 1,679 points, ahead of Claude Fable 5. It also led SWE Marathon and Program Bench across Moonshot's coding suite.

1M token context

Built for long-horizon coding and agent workloads. On BrowseComp, Moonshot notes that using the full 1M window rather than compaction at 300K lifts the score into the 90–91% range.

Native vision

It reads images directly, and independent testing found the image understanding genuinely good — accurate, detailed descriptions from a rendered SVG, not vague summaries.

Kimi Delta Attention

Two architectural additions — Kimi Delta Attention and Attention Residuals — which Moonshot credits for much faster long-context handling and better scaling efficiency than K2.

Open weights

Moonshot committed to releasing full weights for research and commercial use, making K3 the largest open-weight model to date, taking that crown from DeepSeek's 1.6T v4 Pro.

Why use it here

Kimi K3 on Fullmira vs everywhere else

The model is identical. What differs is everything around it.

On Fullmira

  • One subscription covers every model. Kimi K3, GPT-5.6, Claude, Gemini, plus image, video and audio — not a separate bill each.
  • Switch mid-conversation. K3 burning too much reasoning on a simple question? Hand the same thread to a faster model without re-pasting anything.
  • Compare side by side. Send one prompt to K3 and Claude Fable 5 at once — that's where K3 claims the top spot on frontend code.
  • Try before you commit. Sign up free and get 500 credits instantly — no card needed, so you can test a 2.8T model on your own work before setting up an API key.
  • No self-hosting required. Open weights are great, but a 2.8T MoE needs serious hardware. Here it just works in the browser.

Single-vendor apps

  • A separate subscription or API key for each provider you want access to.
  • Locked to one vendor — no switching when another model suits the task better.
  • Comparing means copying your prompt into a different tab by hand.
  • Running the open weights yourself means sourcing hardware for a 2.8T MoE.
  • Your work spread across several apps with separate histories.

When Kimi K3 is the right pick

K3 matters for a specific reason: it's the first Chinese open-weight release that competes with the top American systems on capability rather than on price. That reframes what an open model can be used for.

Frontend and UI code

This is its clearest win. K3 ranks first on Arena.ai's Frontend Code leaderboard at 1,679 points, ahead of Claude Fable 5, in blind developer voting — a 17-place jump from K2.6 with a 76% pairwise win rate.

Long-horizon agentic work

It led SWE Marathon and Program Bench, and trailed GPT-5.6 Sol on Terminal-Bench 2.1 by only half a point. Combined with the 1M-token window, that suits tasks which run for a long time and accumulate context.

The efficiency story

Artificial Analysis put K3's overall Elo at 1,547 — a 732-point jump from K2.6, behind only Claude Fable 5. Estimated cost per task came in at roughly half that of Claude Opus 4.8, and Moonshot reports it uses 21% fewer output tokens than K2.6 on equivalent tasks.

Where Kimi K3 isn't the winner

Moonshot's own benchmarks show K3 losing to Claude Fable 5 and GPT-5.6 Sol overall. And it currently ships with only one reasoning effort level, “max” — it reasons heavily on everything. For quick, cheap questions that's the wrong tool.

FAQ

Kimi K3 questions, answered

What is Kimi K3?
Kimi K3 is Moonshot AI's flagship model, released on July 16, 2026. It's a mixture-of-experts system with 2.8 trillion total parameters — the largest open-weight model released to date — routing each token through just 16 of its 896 experts. It has a 1M-token context window and native image understanding.
Can I use Kimi K3 for free?
Yes — sign up free and get 500 credits instantly, no credit card and no Moonshot API key required. The chat box at the top of this page works right now.
Is Kimi K3 really open source?
It's open-weight rather than open-source in the strict sense — Moonshot committed to releasing the full model weights for research and commercial use. Worth being realistic though: a 2.8T mixture-of-experts model needs substantial hardware to run locally.
How good is Kimi K3 compared to Claude and GPT?
Genuinely competitive, but not top of the frontier. Moonshot's own benchmarks show K3 beating Claude Opus 4.8 and GPT-5.5 while losing to Claude Fable 5 and GPT-5.6 Sol overall. Where it leads outright is frontend code — #1 on Arena.ai's Frontend Code arena at 1,679 points.
What is Kimi K3's context window?
1 million tokens, aimed at long-horizon coding and agent workloads. On BrowseComp, running with the full 1M window instead of context compaction at 300K raises the score into the 90.4–91.2% range.
Why does Kimi K3 use so many tokens?
Because it currently ships with only one reasoning effort level, "max". It thinks hard on everything. In independent testing a simple SVG generation consumed over 13,000 reasoning tokens.
Does Kimi K3 understand images?
Yes, it has native vision. Independent testing found the image understanding genuinely strong — given a rendered SVG it produced detailed, accurate alt text describing specific colours, objects and composition rather than a vague summary.
Do I need a separate subscription for Kimi K3?
No. One Fullmira account covers Kimi K3 alongside GPT-5.6, Claude, Gemini and every other model on the platform.