Kimi K3 Explained: Pricing, Specs, and Why You Can't Self-Host It

Kimi K3, from Moonshot AI, is the largest open-weight AI model that exists right now: 2.8 trillion parameters, released as real downloadable weights on July 27, 2026. "Open-weight" is doing a lot of work in that sentence, though. Moonshot's own deployment guidance for K3 is a supernode with 64 or more accelerators — not a gaming PC, not a single server, a small data center. The weights themselves are a 1.56 terabyte download across 96 files. You can absolutely use Kimi K3 today, for free or for pennies per conversation — just not by downloading it onto anything you own.

What Kimi K3 actually is

A customer told Jake she'd read that Moonshot's new AI was "open source and free" and asked him to "install the good one" on the shop's back-office PC instead of paying for a subscription anywhere. Jake pulled up the Hugging Face page, saw the download button, and got as far as watching the file-size estimate climb past a terabyte before he called Ethan.

"Open-weight doesn't mean open-hardware," Ethan told him. "Moonshot published the actual numbers that come out of the model's training, the actual weights — genuinely, not a locked-down demo. It's just that the numbers describe something the size of a shipping container, and your back office has a desk."

Kimi K3 is Moonshot AI's flagship model, announced July 16, 2026 and made available as downloadable weights on July 27. Moonshot's own tech blog calls it "the world's first open 3T-class model" — the first time anyone has released the actual weights of a model this large for anyone to download, rather than keeping it locked behind an API the way OpenAI, Anthropic, and Google keep their frontier models. It's built for long-horizon coding, knowledge work, and reasoning, and it natively understands images and video alongside text, not as a bolted-on add-on.

Here are the numbers that actually define it, straight from the official model card on Hugging Face:

Two architectural choices are doing the heavy lifting here, and they're worth understanding in plain terms since Moonshot leans on them constantly in its own materials. Kimi Delta Attention (KDA) is a more efficient way of letting the model refer back to earlier parts of a long conversation, built to hold up at the million-token scale without the usual attention mechanism eating disproportionately more compute as context grows. Attention Residuals (AttnRes) lets the model selectively pull information from earlier layers instead of just stacking everything uniformly — think of it as a more deliberate memory across the model's depth, not just its context length. Together with a routing system called Stable LatentMoE that decides which 16 of the 896 experts handle a given token, Moonshot says these changes give K3 roughly 2.5× the scaling efficiency of its predecessor, Kimi K2 — more capability per unit of compute, not just more parameters for their own sake.

Is Kimi K3 actually open source?

Not in the strict, OSI-approved sense — but it's genuinely open-weight and genuinely permissive, which is what most people actually mean when they ask this. The weights are released under a custom Kimi K3 License, published alongside the model on Hugging Face. In plain terms: you can use it, modify it, and deploy it commercially, and you're free for the vast majority of real-world use. Two conditions only kick in at serious scale: a product built on K3 that reaches more than 100 million monthly active users, or more than $20 million in monthly revenue, has to display "Kimi K3" prominently; and a business reselling K3 as a hosted API service earning over $20 million a year needs a separate licensing conversation with Moonshot. If you're an individual, a small team, or a mid-sized company, neither of those thresholds is something you'll bump into by accident.

Can you actually run Kimi K3 locally?

This is the question that actually matters, and the honest answer is no — not on anything a person or a small business would describe as "their own hardware." The official Hugging Face repository is 96 safetensors shards totaling roughly 1.56 terabytes, and that's already the compressed, MXFP4-quantized deployment weights, not a bulkier full-precision version you could shrink further. Moonshot's own tech blog states its deployment recommendation directly: "we recommend deploying Kimi K3 on supernode configurations with 64 or more accelerators." That's Moonshot's own guidance for its own model, not a third party's guess.

The Mixture-of-Experts design doesn't rescue this the way people sometimes assume. Yes, only 16 of the 896 experts activate for any given token — but different tokens route to different experts, so a real server still has to keep every expert loaded and ready. Sparsity here cuts the compute needed per token; it does nothing to shrink the storage the model needs to sit in memory. A realistic self-hosted deployment means a multi-node GPU cluster with something in the neighborhood of 1.6 terabytes of aggregate accelerator memory — several 8-GPU servers wired together with high-bandwidth interconnects, not one machine, however good.

What Kimi K3 actually costs

Through Moonshot's own API platform, Kimi K3 is priced per million tokens, with a meaningful discount for cache hits:

• $0.30 per million input tokens on a cache hit
• $3.00 per million input tokens on a cache miss
• $15.00 per million output tokens

That cache-hit discount matters more than it might look at first glance: Moonshot states its infrastructure, built on something it calls Mooncake, achieves a cache hit rate above 90% on typical coding workloads. In practice, that means most of a real coding session's input tokens are being billed at the $0.30 rate, not $3.00 — the effective cost of a long, back-and-forth coding session is usually much closer to the cheap end of that range than the sticker price alone suggests.

Kimi K3 vs. Claude, DeepSeek, and Qwen3.8

Moonshot's own tech blog is unusually candid about where K3 sits: it states plainly that K3's overall performance "still trails the most powerful proprietary models, Claude Fable 5 and GPT 5.6 Sol," while "consistently outperforming other tested models" like GLM-5.2. That's a genuinely honest positioning statement from a model's own maker, and it's a good starting point for a real comparison:

Worth flagging honestly: DeepSeek's pricing has changed several times through 2026 — new model names, a peak/off-peak billing split added in August, and at least one deprecation announcement that didn't actually take effect on schedule. Treat the DeepSeek row as a snapshot, not a fixed number, and check its live pricing page before budgeting around it. Kimi K3, Qwen3.8 Max, and Claude Fable 5.1's numbers are more stable and corroborated across several independent trackers as of this writing.

The pattern that actually matters: Kimi K3 undercuts Claude Fable 5.1 by roughly 15–30× depending on cache hits, while landing in a similar price range to Qwen3.8 Max — but with genuinely open weights behind it, which Qwen's flagship API tier doesn't offer at all.

What Kimi K3 is actually good at

Moonshot's own case studies lean heavily on long-horizon, minimally-supervised work — tasks where a model has to keep making sound decisions across hours of work, not just answer one prompt well. A few of the more concrete examples from the tech blog: K3 built a compact Triton-like GPU compiler called MiniTriton from scratch, complete with its own intermediate representation and code-generation pipeline, and Moonshot reports it matching or beating existing optimized compilers on several benchmarks. In another test, given 48 autonomous hours and open-source chip-design tools, K3 designed, optimized, and verified a small chip that closes timing at 100 MHz. And on a research-reproduction task — recreating a set of universal relations in computational astrophysics from 20-plus source papers — Moonshot says K3 finished in about two hours what typically takes an experienced researcher one to two weeks.

On the benchmark side, one number worth noting on its own: on BrowseComp, an agentic web-browsing benchmark, K3 scores 90.4 when given its full 1-million-token context with no compaction — a case where that huge context window is doing real, measurable work, not just existing as a marketing spec.

The limitations Moonshot admits to itself

Model announcements rarely list their own weaknesses this directly, so it's worth repeating Moonshot's three stated limitations rather than letting them get lost:

1. It's sensitive to how "thinking history" is passed back to it. K3 was trained expecting its own prior reasoning to be preserved and resupplied on each turn. If the tool you're using doesn't do that correctly — or if you switch mid-conversation from a different model to K3 — output quality can become unpredictable. Moonshot recommends a harness with verified compatibility, like Kimi Code, and starting a fresh session rather than switching models mid-stream.
2. It can be excessively proactive on long tasks. Because K3 is trained to push through ambiguity on long-horizon work, it may make decisions on your behalf when it hits a minor snag or an unclear instruction, rather than stopping to ask. If your use case needs the model to stay strictly inside defined boundaries, Moonshot's own advice is to spell that out explicitly in the system prompt or an AGENTS.md file, not to assume restraint by default.
3. There's a real user-experience gap against Claude Fable 5 and GPT 5.6 Sol, despite competitive benchmark scores — Moonshot's own words, not a hedge added elsewhere. Benchmarks and day-to-day feel aren't the same thing, and this model's maker says so directly.

How to actually use Kimi K3 today

Four real, documented ways to reach K3, none of which require the hardware conversation above:

1. The Kimi app — iOS, Android, and HarmonyOS, or directly at kimi.ai. This is the free, conversational way in.
2. Kimi Work — a desktop app (version 3.1.0 or later, Windows and Apple Silicon Macs) for document, spreadsheet, slide, and research-style work, with K3 behind it.
3. Kimi Code — a terminal-based coding agent. Run it and select K3 with the /model command once it's active.
4. The Kimi API — go to platform.kimi.ai, and specify the model ID kimi-k3 in your requests, at the per-token pricing above.

Frequently asked questions

What is Kimi K3?

Kimi K3 is Moonshot AI's flagship model, a 2.8-trillion-parameter open-weight Mixture-of-Experts model with native vision and a 1-million-token context window, released July 27, 2026. It's designed for long-horizon coding, agentic knowledge work, and reasoning.

Is Kimi K3 free to use?

Yes, through the Kimi app and Kimi Work desktop app at no cost for individual use. The API is metered per token ($0.30–$3.00 input, $15.00 output per million), and self-hosting the open weights avoids per-token fees but requires a real GPU cluster, not free hardware you already own.

Is Kimi K3 open source?

It's open-weight under a custom Kimi K3 License, not an OSI-approved open-source license. In practice it's commercially permissive for the vast majority of use: attribution requirements only kick in above 100 million monthly users or $20 million in monthly revenue.

How much does the Kimi K3 API cost?

$0.30 per million input tokens on a cache hit, $3.00 per million on a cache miss, and $15.00 per million output tokens, per Moonshot's official API pricing. Cache hit rates above 90% on coding workloads mean real-world costs usually sit closer to the cheaper end.

Can Kimi K3 be run locally?

Not on consumer or even typical workstation hardware. The weights are 1.56 terabytes across 96 files, and Moonshot's own guidance recommends supernode configurations with 64 or more accelerators. Realistic self-hosting needs a multi-node cluster with roughly 1.6 terabytes of aggregate GPU memory.

What are Kimi K3's hardware requirements to self-host?

A multi-node GPU cluster, not a single machine. The Mixture-of-Experts design reduces per-token compute, not storage — every one of the 896 experts still has to be loaded, since different tokens can route to different experts.

What is Kimi K3's context window?

1,048,576 tokens — roughly 1 million. On the BrowseComp agentic benchmark, Moonshot reports a score of 90.4 when the full window is used with no context compaction.

Does Kimi K3 have vision (image) capabilities?

Yes, natively. It uses a dedicated vision encoder called MoonViT-V2 with 401 million parameters, built into the model rather than added as a separate system.

When was Kimi K3 released?

It was announced on July 16, 2026, with the full open weights, inference code, and technical report released on July 27, 2026 (July 28 in China Standard Time).

How many parameters does Kimi K3 have, and how many are active?

2.8 trillion total parameters, with roughly 104 billion active per token through its sparse Mixture-of-Experts design — 16 of 896 experts selected per token, plus 2 shared experts.

How does Kimi K3 compare to Claude Fable 5?

Moonshot's own materials state K3's overall performance still trails Claude Fable 5, while remaining frontier-level and undercutting it by roughly 15–30x on API pricing. Fable 5.1 is API-only with no open weights at all.

How does Kimi K3 compare to DeepSeek V4?

DeepSeek's models are considerably smaller and more genuinely self-hostable, and typically cheaper per token, though DeepSeek's pricing has shifted several times in 2026 and is worth checking directly before comparing. Kimi K3 is the larger, less accessible-to-self-host model of the two.

How does Kimi K3 compare to Qwen3.8 Max?

They land in a similar API price range ($2.00/$6.00 for Qwen3.8 Max vs. Kimi K3's $0.30–$3.00/$15.00), but Kimi K3 has fully open weights while Qwen3.8 Max's flagship API tier does not — Alibaba's comparable open checkpoint is a separate, smaller release.

Where can I download Kimi K3's weights?

The official repository is on Hugging Face under moonshotai/Kimi-K3, with code and the technical report also on Moonshot AI's GitHub. Expect a roughly 1.56 terabyte download across 96 shard files.

What are Kimi K3's known limitations?

Moonshot states three directly: sensitivity to how prior reasoning history is passed back to it, a tendency toward excessive proactiveness on long tasks without explicit constraints, and a noticeable user-experience gap against Claude Fable 5 and GPT 5.6 Sol despite competitive benchmarks.

Can I use Kimi K3 with Claude Code or other coding tools?

Moonshot recommends Kimi Code specifically, since K3 was trained expecting its own thinking-history format to be preserved correctly, and Moonshot states this is best guaranteed with a harness it has verified. Mid-session switches from another model to K3 are explicitly discouraged.

Revision note. Written September 2026. Pricing (especially DeepSeek's, which has moved several times this year already) and the self-hosting tooling around K3 are the parts most likely to change next — check Moonshot's own API pricing page and the Hugging Face repository directly before treating any number here as fixed. If you came here after almost downloading 1.56 terabytes onto a machine that couldn't have run it, you're in good company — that mismatch between "open" and "runnable" catches serious engineers too, not just first-timers.