Two days after Moonshot AI released Kimi K3 to the public, the Beijing startup had to pause new subscriptions. Demand had outstripped its compute capacity. Within the same window, shares of rival Chinese AI firms Z.ai and MiniMax Group fell as much as 30% and 16% respectively, Alibaba dipped, and Moonshot found itself fielding accusations from the White House itself. That is an unusual amount of turbulence for a single model release, and it’s why “Kimi AI with K3” has become a genuine talking point on both sides of the Pacific rather than just another entry on a leaderboard.
Here’s what Kimi K3 actually is, what it can do, what it costs, and why a Chinese open-weight model has managed to rattle Wall Street, Silicon Valley, and Washington all at once.
What Is Kimi K3?
Kimi K3 is the flagship large language model from Moonshot AI, a Beijing-based lab that has been building the Kimi model family since 2023. Moonshot unveiled K3 on July 16, 2026, positioning it as the successor to Kimi K2.6 and K2.7 Code, and followed up on July 27 by publishing the full model weights on Hugging Face — making it, by Moonshot’s own framing, the largest openly downloadable AI model in the world.
At the center of the launch is a genuinely large number: 2.8 trillion total parameters, organized as a sparse Mixture-of-Experts (MoE) network that activates only 16 of 896 experts for any given token. That sparsity is what keeps inference costs manageable despite the model’s overall size — you’re paying for a fraction of that 2.8 trillion at any one moment, not all of it.
K3 ships in two variants:
- K3 Max — the general-purpose model for chat, coding, and single-agent tasks.
- K3 Swarm Max — a version tuned for large-scale, multi-agent workloads, built around Moonshot’s “Agent Swarm” orchestration system.
Architecture and Technical Specs
| Spec | Detail |
|---|---|
| Total parameters | ~2.8 trillion (MoE) |
| Active parameters per token | 16 of 896 experts (Stable LatentMoE) |
| Context window | 1,048,576 tokens (1M) |
| Attention mechanism | Kimi Delta Attention (KDA) + Attention Residuals |
| Vision | Native, via MoonViT-3D encoder, trained in from pretraining (not bolted on) |
| Reasoning | Always-on “thinking mode”; configurable effort (low / high / max) |
| Agent orchestration | Agent Swarm — up to 300 parallel sub-agents, 4,000+ coordinated tool calls |
| License | Bespoke “Kimi K3 License,” MIT-inspired |
| Download size | Roughly 594GB (native MXFP4) |
The two technical pieces Moonshot is leaning on hardest are Kimi Delta Attention and Agent Swarm. KDA maintains a fixed-size learned state instead of letting attention cost scale with the full context length, which Moonshot says makes inference at the full 1-million-token window up to six times faster and cheaper than a conventional approach. Agent Swarm, meanwhile, lets K3 spin up to 300 sub-agents that work a problem in parallel — research, coding, browsing, and visual analysis happening simultaneously rather than sequentially — which Moonshot claims delivers roughly a 4.5x speedup on wide-search and long-horizon tasks compared to a single agent working alone.
Vision is also worth flagging as a real capability rather than a marketing checkbox: because MoonViT-3D was trained into the model from the start rather than added afterward, every sub-agent in a swarm can independently read screenshots, charts, and UI mockups without routing through a separate vision model.
Key Features
Kimi Code — Moonshot’s terminal-based coding agent, comparable in role to Claude Code or Codex CLI. K3 is selectable inside it via a /model command, and Moonshot explicitly recommends Kimi Code as K3’s primary agent harness.
Kimi Work — A desktop knowledge-work environment (version 3.1.0+ required for K3) that bundles chat, document analysis, and agentic task execution into one app.
Visual Agentic Slides — Upload reference materials and K3 will generate presentation decks matched to that visual style, aimed at cutting out manual deck formatting.
AAA Game Generation — Perhaps the flashiest feature Moonshot is marketing: a single prompt can produce a playable multiplayer 3D game, with K3 handling world-building, character design, levels, and combat logic, then iterating on its own playtesting feedback.
Always-on reasoning — Unlike some competitors that let you toggle reasoning off for speed, K3 has no non-thinking mode. Every response includes a reasoning trace, and that trace is billed at the same rate as the visible output — worth knowing before you assume a “quick” query will be cheap.
Tiered context length — On the API, the full 1M-token window is available flat, with no long-context surcharge. Inside the consumer apps, though, the extra-long context is gated to the top two membership tiers — an unusual choice that effectively turns context window size into a subscription perk rather than a model constant.
How Kimi K3 Stacks Up: Benchmarks
Moonshot published a substantial benchmark suite alongside the launch, and independent trackers like Artificial Analysis and LMArena have since run their own comparisons against Anthropic’s Claude Fable 5 and Claude Opus 4.8, and OpenAI’s GPT-5.6 Sol.
| Benchmark | Kimi K3 | Claude Fable 5 | GPT-5.6 Sol | Claude Opus 4.8 |
|---|---|---|---|---|
| Artificial Analysis Intelligence Index | 57 | 60 | 59 | 56 |
| Frontend Code Arena (LMArena) | 1,679 (#1) | 1,631 | — | — |
| SWE Marathon (long-horizon coding) | 42.0 | 35.0 | 39.0 | 40.0 |
| Terminal-Bench 2.1 | 88.3 | 84.6 | 88.8 | — |
| Program Bench | 77.8 | 76.8 | 77.6 | — |
| GPQA Diamond | 93.5% (best open-weight) | — | — | — |
| GDPval v2 (agentic, Elo) | 1,668 | 1,760 | 1,494* | 1,600 |
| CharXiv (visual agent) | 91.3 | 93.5 | — | — |
*GPT-5.5 figure as reported in the source dataset.
The pattern that emerges: K3 doesn’t beat the top US closed models across the board, but it wins decisively on coding tasks that run long — front-end code generation, terminal-based engineering sessions, and multi-hour “marathon” software work. It lands a close fourth on general intelligence and trails the leaders on broad agentic and visual-agent evaluations. For an open-weight model, taking the #1 spot on LMArena’s Frontend Code Arena — a first for any openly downloadable model — is a genuinely notable result, not just a self-reported flex.
A caveat worth stating plainly: several of the headline scores (including Kimi Code Bench 2.0) come from Moonshot’s own release materials, and different agent scaffolds were used across the comparisons — meaning these aren’t perfectly clean model-to-model numbers. Treat them as a directional map of K3’s strengths rather than a definitive ranking.
Pricing: API and Membership
Kimi K3 is priced well above the “ultra-cheap Chinese model” reputation that DeepSeek and earlier Kimi releases built. It’s positioned closer to a mid-tier Western model than a budget option.
API pricing:
| Item | Price |
|---|---|
| Input tokens | $3.00 / million |
| Output tokens (incl. reasoning) | $15.00 / million |
| Cached input tokens | $0.30 / million (90% discount) |
| Context surcharge | None — flat across the full 1M window |
For reference, that’s roughly in line with Claude Sonnet 5’s API pricing, and it undercuts GPT-5.6 Sol and Claude Opus 4.8 on output cost while landing well above DeepSeek V4 ($0.87/M output) and z.ai’s GLM-5.2 ($4.40/M output). Because K3 gets high cache-hit rates on coding workloads, Artificial Analysis estimates a blended effective rate closer to $2.31 per million tokens in practice.
Consumer membership (kimi.com, Kimi Work, Kimi Code):
| Tier | Monthly price | Notes |
|---|---|---|
| Free | $0 | Rate-limited, no API access, K3 included with usage caps |
| Moderato | $19 | Entry paid tier |
| Allegretto | $39 | Mid tier |
| Allegro | $99 | Unlocks full 1M-token context; comparable Kimi Code quota to Anthropic’s $200/month plan, per early user reports |
| Vivace | $199 ($159/mo billed annually) | Highest quota, full context, fastest concurrency |
A few things worth flagging before you build a workflow around these numbers. First, membership and API billing are entirely separate — a subscription doesn’t hand you API credits. Second, Moonshot’s pricing page has carried a notice that Kimi and Kimi Code benefits may be split into separate subscription products going forward, so treat current tiers as a snapshot rather than a fixed structure. Third, only the top two tiers get the full 1-million-token context; everything below that runs on a shorter window despite running the same underlying model.
The “Open Weight” License Has a Catch
Moonshot is marketing K3’s July 27 weight release as proof of openness, and technically, you can download, self-host, fine-tune, and modify the model for free. But the “Kimi K3 License” is a bespoke document, not MIT or Apache 2.0, and Hugging Face tags it as “license:other.”
The terms that matter for anyone thinking about building a business on it:
- Model-as-a-Service threshold: If you or your affiliated companies run inference-as-a-service on K3 and clear $20 million in revenue over any trailing 12 months, you’re required to sign a separate commercial agreement with Moonshot before continuing.
- Attribution requirement: Products built on K3 that exceed 100 million monthly active users or $20 million in monthly revenue must visibly attribute Kimi K3 in the product.
- Internal-use carve-out: None of this applies if you’re running K3 purely inside your own organization, or accessing it through Moonshot’s official products or certified partners.
It’s a genuinely permissive license for internal use, hobbyists, and small products — but it’s not the anything-goes MIT license the “Modified MIT” label on earlier Kimi releases implied. It’s closer in spirit to Meta’s Llama community license, which gates commercial rights by scale rather than by use case.
There’s also a practical hosting reality worth knowing: at roughly 594GB and built on custom components (including KDA and a nonstandard expert-routing design), K3 isn’t yet supported in mainstream local-inference tools like Ollama or llama.cpp, and running it at full fidelity requires something like eight H100-class GPUs. For most teams, “open weight” currently means “hosted by a third party,” not “runs on my laptop.”
Why Kimi K3 Became a US Headline
This is the part that’s made Kimi K3 more than a product story.
The market reaction was immediate. In the days after launch, competing Chinese AI stocks took a real hit — Z.ai fell as much as 30% in Hong Kong trading and MiniMax Group dropped as much as 16%, while Alibaba slipped too. Moonshot’s own numbers moved the other direction: daily revenue reportedly grew at least sixfold after K3 shipped, annual recurring revenue hit $300 million in June (up from $200 million in April), and the company is reportedly raising a new round at a $50 billion valuation ahead of a possible Hong Kong IPO. Demand actually outran Moonshot’s own infrastructure — new paid subscriptions were briefly suspended just two days after launch.
US companies are already building on it. Cursor has used Kimi models to help build its Composer 2 coding agent. DoorDash CTO Andy Fang has said the company routes lower-level engineering work to Kimi K2.6. Thinking Machines used an earlier Kimi model to help generate post-training data for its Inkling model. More broadly, Chinese open-weight models are now reported to account for a majority of Chinese-model token usage on routing platforms like OpenRouter among US developers — meaning the “just use the free Chinese model” pattern that started with DeepSeek in early 2025 is continuing, not fading.
Washington’s response has been the sharpest part of the story. Six days after K3’s July 16 debut, Michael Kratsios, director of the White House Office of Science and Technology Policy, publicly accused Moonshot of running a large-scale, covert distillation operation against Anthropic’s Fable 5 model to help build K3, and alleged the company had also accessed export-restricted Nvidia chips. Treasury Secretary Scott Bessent said the administration would consider sanctions and Entity List measures if the distillation allegations rose to the level of IP theft, and Undersecretary of State Jacob Helberg went further, describing the situation as a theft of American intellectual property.
The accusation didn’t come out of nowhere. In February 2026, Anthropic had already gone public with claims that Moonshot, DeepSeek, and MiniMax ran coordinated distillation campaigns against Claude, attributing roughly 24,000 fraudulent accounts and more than 16 million exchanges to the three companies combined, with Moonshot alone linked to 3.4 million of those interactions. After K3 launched, screenshots circulated online showing the new model identifying itself as “Claude, an AI assistant made by Anthropic” when asked its name — the kind of artifact that tends to happen when a model has been trained on outputs from another system, though it isn’t conclusive proof on its own.
Moonshot has denied the allegations. Its head of enterprise business, Huang Zhenxin, rejected the distillation claims directly in an interview with Chinese state media. Independent analysts have also expressed skepticism about the specific timeline the White House described — Fable 5 was only reinstated on July 1 after a brief export-control-related suspension, which some researchers argue left Moonshot a narrow window to have distilled meaningfully from it before K3 shipped on July 16. Other researchers have noted that K3’s output patterns are broadly consistent with heavy distillation of coding data more generally, without necessarily pinning that to Fable 5 specifically.
Whatever the eventual conclusion on distillation, the bigger structural story is one that first surfaced with DeepSeek in early 2025 and is now recurring: a Chinese lab has again produced a model that competes with the US frontier on real benchmarks, at a fraction of the presumed cost, using chips the US has spent years trying to keep out of Chinese hands. That’s renewing debate in Washington over whether export controls are actually working as intended, and reports suggest officials are now discussing whether to restrict access to Chinese open-weight models altogether.
Reality Check: What the Marketing Doesn’t Tell You
In keeping with an honest look at the hype rather than just repeating Moonshot’s press materials:
- “Cheapest frontier model” is no longer accurate. K3’s API pricing ($3/$15 per million tokens) is roughly Claude Sonnet territory — a real jump from K2’s budget positioning, and well above DeepSeek V4 or GLM-5.2. If you came to Kimi expecting DeepSeek-level pricing, K3 isn’t that model anymore.
- “Open weight” comes with real strings attached. The $20M revenue trigger and attribution rules mean K3 isn’t a no-conditions-attached open-source release, whatever the “Modified MIT” branding on earlier Kimi models implied.
- Self-hosting is aspirational for most teams right now. Custom architecture, a ~594GB download, and a minimum of eight H100-class GPUs put genuine self-hosting out of reach for anyone without serious infrastructure, at least until mainstream tooling catches up.
- Benchmark wins are real but scaffold-dependent. K3’s Frontend Code Arena #1 finish and SWE Marathon lead are independently verifiable through LMArena; several of the other headline numbers are Moonshot’s own and were run with different agent harnesses across comparisons, which limits how cleanly you can compare them model-to-model.
- The distillation story is unresolved. As of this writing, the White House has made the allegation publicly but has not released supporting evidence, and Moonshot has denied it. Readers should treat this as an open dispute, not a settled fact in either direction.
FAQ
Is Kimi K3 free to use? There’s a free tier at kimi.com with rate limits and no API access. Full API and higher-tier membership access are paid.
Is Kimi K3 better than Claude or GPT-5.6? It depends on the task. K3 leads on front-end coding and long-horizon software engineering benchmarks. Claude Fable 5 and GPT-5.6 Sol still lead on general intelligence and broad agentic evaluations.
Can I run Kimi K3 on my own hardware? Yes, in principle — the weights are downloadable — but practically it requires multi-GPU server infrastructure (roughly eight H100-class GPUs) and tooling support that’s still catching up as of this writing.
Why is the US government concerned about Kimi K3? White House officials have publicly alleged Moonshot used outputs from Anthropic’s Fable 5 model to help train K3 and accessed export-restricted chips. Moonshot denies both claims. The dispute is feeding a broader debate about whether US export controls are effectively slowing China’s AI progress.
What does “K3 Swarm Max” mean? It’s the variant of K3 optimized for Agent Swarm workloads — coordinating up to 300 sub-agents on a single task rather than running one agent sequentially.
Bottom Line
Kimi K3 is a legitimately capable model — the coding benchmarks aren’t vapor, and the price-to-performance ratio on long-horizon engineering work is competitive with anything US labs currently ship. But it arrives wrapped in more controversy than capability alone would explain: a licensing scheme that’s less open than it sounds, a pricing story that no longer matches the “cheap Chinese AI” narrative, and an unresolved dispute with the White House and Anthropic over exactly how it was built. Whether K3 ends up remembered as a genuine engineering achievement or as another data point in an unresolved distillation dispute will likely depend on evidence nobody involved has published yet.
