Kimi K3: China's Moonshot Ships the Largest Open AI Model

Moonshot AI released Kimi K3 on July 16, 2026, and it's now the largest open-weight AI model ever shipped — a 2.8 trillion-parameter model that lands close to GPT-5.6 Sol and Claude Fable 5 on independent benchmarks, at a fraction of their API price. It's the clearest sign yet that the gap between "open" and "frontier" AI models has nearly closed, and it's happening faster on price than anyone expected.

If you build products on top of AI APIs, choose models for a startup budget, or just want to know whether the cheap open alternatives are finally good enough to trust, Kimi K3 is worth a look. Here's what it actually is, how it compares to the alternatives, and where the catch is.

Quick Take

• What it is: A 2.8 trillion-parameter Mixture-of-Experts model from Chinese lab Moonshot AI, with a 1-million-token context window and native image understanding.
• Why it matters: It's the largest open-weight model released to date — about 75% bigger than DeepSeek's V4 Pro (1.6T parameters) — and it scores close to, though still behind, the best closed models from OpenAI and Anthropic on independent benchmarks.
• Pricing: $0.30 per million cached input tokens, $3 per million uncached input tokens, $15 per million output tokens — roughly a third of GPT-5.6 Sol's list price.
• The catch: It shipped via Moonshot's app and API on July 16, but the actual downloadable weights aren't out until July 27, 2026. Until then, "open" is a promise, not something you can self-host.

What Kimi K3 Actually Is

Kimi K3 activates just 16 of its 896 "experts" per request, so despite the eye-catching 2.8 trillion total parameter count, it runs closer to a ~50-billion-parameter model in practice — that's the point of a Mixture-of-Experts architecture. Moonshot shipped two variants at launch: K3 Max for general chat and coding agents, and K3 Swarm Max, tuned for running many parallel agent tasks at once.

It's available now inside the Kimi app, through Moonshot's Kimi Code tool, via Moonshot's own API, and through third-party routers like OpenRouter — so you don't need a direct account with a Chinese vendor to try it. Full open weights, under a modified MIT license, are scheduled for July 27, 2026, roughly two weeks after the initial launch.

How It Actually Stacks Up

Moonshot's own announcement claims K3 "substantially outperformed" Claude's Opus 4.8 and OpenAI's GPT-5.6 Sol, and performs competitively with Claude Fable 5, Anthropic's current flagship. Independent testing tells a more modest story. Artificial Analysis, a widely cited third-party benchmark tracker, scored K3 Max at 57.1 on its Intelligence Index (v4.1) — just behind GPT-5.6 Sol Max at 58.9, and behind Fable 5 Max at 59.9. That's a real gap, not a rounding error, though it's a much smaller gap than open models have posted in the past.

Model	Type	Context window	Output price
Kimi K3 Max	Open (weights July 27)	1M tokens	$15 / M tokens
DeepSeek V4 Pro	Open	1M tokens	Peak/off-peak pricing
GPT-5.6 Sol	Closed	—	$30 / M tokens

Pricing verified July 2026 from vendor-published rates; confirm current rates directly with each provider before committing.

The closest open competitor is DeepSeek's V4 Pro, a 1.6-trillion-parameter model that also went fully official in mid-July 2026. DeepSeek's twist is different: instead of cutting prices further, it introduced peak-time surcharging — API costs double during business hours (9am–12pm and 2pm–6pm) and stay at the old rate off-peak. Kimi K3 doesn't do that; its pricing is flat, which makes cost genuinely easier to predict for anyone building a product on top of it.

The Catch: "Open" Isn't Available Yet

The most important asterisk on this launch is timing. Kimi K3 is usable today only through Moonshot's hosted app and API — the actual model weights, the thing that lets you run it on your own infrastructure, don't land until July 27. Until then, calling it "open" is technically accurate but practically premature; you're renting access to it exactly like you would GPT-5.6 or Claude, just at a much lower price.

There's also a real business consideration that coverage of this launch has mostly glossed over: routing production traffic, especially anything involving customer data, through a Chinese-hosted API carries data-residency and compliance implications that vary by industry and region. Using Kimi K3 through a US-based router like OpenRouter sidesteps some of that, but it doesn't erase the underlying question — worth a real conversation with legal or compliance before it touches anything sensitive.

Should You Actually Use It?

For side projects, internal tools, or any workload where you're paying per-token and watching margins closely, Kimi K3's pricing is hard to ignore — output tokens at $15/M versus GPT-5.6 Sol's $30/M is a meaningful difference at scale, and it doesn't come with DeepSeek's peak-hour surcharge. For coding agents and long-document workloads, the 1M-token context window puts it in the same tier as the current top models.

For anything customer-facing where the last few points of reasoning quality matter, or where compliance requirements are strict, the independent benchmarks say GPT-5.6 Sol and Claude Fable 5 still hold a real, if narrowing, edge — a gap we've tracked as these three families have leapfrogged each other all year in our Claude vs. ChatGPT vs. Gemini comparison. If you're building AI-assisted developer tools rather than chat products, it's also worth weighing against the agentic coding tools we cover in Cursor vs. GitHub Copilot vs. Claude Code, several of which already support routing to alternative model backends like Kimi.

Frequently Asked Questions

Is Kimi K3 actually open source? Not yet in the fullest sense. It's usable via API and app today, but the downloadable weights — what most people mean by "open" — are scheduled for July 27, 2026, under a modified MIT license.

Is Kimi K3 better than GPT-5.6 or Claude? Moonshot claims it beats both on internal testing. Independent scoring from Artificial Analysis puts it close behind GPT-5.6 Sol and Claude Fable 5, not ahead of them — a near-frontier result, not a frontier-beating one.

How much does Kimi K3 cost to use? $0.30 per million cached input tokens, $3 per million uncached input tokens, and $15 per million output tokens through Moonshot's own API, verified July 2026.

Can I access it without a Chinese vendor account? Yes — it's listed on third-party routers like OpenRouter, which lets you call it through a US-based intermediary rather than signing up directly with Moonshot.

Does this affect DeepSeek's position? Somewhat. DeepSeek's V4 Pro remains the more established open option, but its new peak-time pricing makes costs less predictable than Kimi K3's flat rate, even though DeepSeek's model is smaller and cheaper off-peak.

Bottom Line

Kimi K3 doesn't dethrone GPT-5.6 or Claude, but it proves the price-to-capability gap between open and closed AI models is closing fast. For cost-sensitive, high-volume workloads it's a legitimate option worth testing now through a router like OpenRouter; for anything where reasoning quality or compliance is non-negotiable, wait and watch how it performs once the actual weights ship on July 27. For more on picking the right model for your budget, see our breakdown of AI tools actually worth paying for in 2026, or browse our full AI Tools coverage.