Kimi K3, by the numbers, as of July 17, 2026:
2.8 trillion parameters. The largest open-weight model ever built, 75% bigger than DeepSeek V4 Pro.
896 experts inside the architecture. 16 fire per token.
1,000,000 tokens of context. Native vision. One reasoning mode, permanently set to max.
$3 per million input tokens, $15 per million output. Cached input drops to $0.30, and Moonshot's serving stack keeps cache hit rates above 90% in coding sessions.
On Arena's independent front-end coding tests, K3 beat Claude Fable 5 and GPT-5.6 Sol. In Arena's broader text ranking, it finished ahead of Opus 4.8 while costing 40% less per task.
Full weights drop July 27 under a Modified MIT license. First open 3T-class model in history.
Moonshot AI, the Beijing lab behind it, passed $200 million in annualized revenue back in April. On July 16 they released K3 and markets had their second DeepSeek moment in 18 months.
Those are the numbers. Now the story behind them, because the numbers alone undersell how strange this release is.
A comeback nobody scheduled
Eighteen months ago Moonshot looked finished. DeepSeek ate their consumer market, their enterprise story stalled, and the Kimi brand read like a footnote in China's AI race. Founder Yang Zhilin, a former Google researcher, kept the lab pointed at one thing: agentic coding models with absurd context windows.
K2 arrived in July 2025 as a solid open-weight coder. K2.5 and K2.6 followed through spring 2026, and by April, Artificial Analysis ranked K2.6 as the strongest open-weight model on their intelligence index. Respectable. Still a tier below the closed frontier.
K3 closed that tier. Moonshot timed the launch days before the World Artificial Intelligence Conference in Shanghai, and the message underneath the benchmarks was blunt: three years of GPU export controls did not stop a mid-size Beijing lab from reaching the frontier and then handing the weights to anyone with a download link.
Anthropic has accused Moonshot and other Chinese labs of industrial-scale distillation, allegedly training on millions of exchanges with American frontier models. Moonshot disputes this. Both things can matter at once: the provenance fight is real, and so is the artifact sitting on Hugging Face in 10 days.
What 2.8 trillion parameters actually buys
The headline number misleads if you read it as raw bulk. K3 is a sparse Mixture-of-Experts model: 896 specialist sub-networks, 16 activated per token. You get the knowledge capacity of a 2.8T model with the inference cost of something far smaller.
Two internal inventions carry the design. Kimi Delta Attention, a hybrid linear attention mechanism, is why the 1M context window exists at a price you can stomach. Attention Residuals, a drop-in replacement for standard residual connections, is where Moonshot claims consistent scaling gains. Both were published as open research on GitHub before the model shipped, which bought K3 credibility with researchers before a single benchmark landed.
The practical translation: this model reads an entire codebase, a year of documents, or 50 video transcripts in one prompt, holds all of it in working attention, and reasons across the whole thing. RAG pipelines, chunking strategies, embedding databases, the entire retrieval industry built to compensate for small context windows, all of it becomes optional for a growing class of tasks.
Add native vision and the input surface widens again. Screenshots, diagrams, whiteboard photos, charts. K3's Arena wins came in front-end coding specifically, the exact discipline where seeing a design and writing the code for it live in the same skull.
The economics are the real weapon
Set the benchmarks aside. The pricing table is where K3 does damage.
$3 in, $15 out puts K3 at the top of Chinese lab pricing and roughly half the per-task cost of Opus 4.8. Then caching rewrites the math. At $0.30 per million cached input tokens with 90%+ hit rates in long coding sessions, the effective input cost of an agent that keeps rereading the same repository collapses by about 4x.
Long-horizon agents live and die on this. An agent that grinds through a repository for 6 hours rereads the same context thousands of times. Under most pricing models that loop bankrupts you. Under K3's cache economics it costs lunch money.
Moonshot also claims K3 spends 21% fewer output tokens than K2.6 on equivalent tasks. Their number, from their evaluation table, so hold it loosely. Independent testers found the opposite pressure at the small end: the always-on max reasoning mode burned 13,241 thinking tokens on a trivial SVG drawing, about $0.25 for one throwaway query. K3 has no economy gear. You cannot ask it to think less.
Which draws the honest usage line. Simple, high-volume, latency-sensitive tasks are the wrong home for this model. Long sessions over massive context, where cache absorbs the input cost and the task justifies max reasoning, are where the pricing turns from expensive to unfair.
July 27 changes the category
Until now, one rule held across the industry: frontier capability lives behind an API. You rent it, the vendor can reprice it, deprecate it, or quietly change its behavior, and your business absorbs whatever happens.
On July 27 the K3 weights land under Modified MIT. Download once and no lab on earth can take the capability back. Fine-tune it on your domain. Run it air-gapped. Serve it from your own metal. Governments, hospitals, banks, and every founder who lost sleep over model deprecation notices now has a frontier-class fallback that answers to nobody.
Almost nobody will self-host 2.8 trillion parameters. The hardware bill for serving a model this size, even sparse, sits far beyond hobbyist range. That misses the point. The weights existing in public permanently caps what anyone can charge for closed models of similar strength, and it guarantees a market of cheap third-party hosts competing to serve K3 at commodity margins. You benefit from the open release even if you never download a single shard.
What to build with it this month
A model with 1M context, native vision, frontier coding scores, and collapsing cache costs is not a chat upgrade. It rewards a different shape of work.
Feed it whole things. Entire repositories for review, complete contract folders for audit, a full competitor content library for teardown. Anything you used to chop into chunks, stop chopping.
Run it long. K3's official positioning is long engineering sessions with minimal supervision: navigate the repo, orchestrate terminal tools, keep going. Queue real multi-hour tasks in the evening and inspect finished work in the morning, with cache eating the cost of every reread.
Point the camera at problems. Screenshot a competitor's landing page and ask for the rebuild. Photograph the whiteboard and ask for the implementation. Vision plus frontier front-end scores makes screen-to-code the model's home turf.
And keep one hand on the meter. Route your trivial, high-frequency calls to a cheap small model, because K3 will happily spend a quarter thinking hard about nothing.
The frontier used to be a subscription. In 10 days it becomes a file. Plan like it.
Thanks for reading this far.
I break down AI models, agent workflows, and the systems behind them, with real numbers and honest caveats. If this was useful, a follow means the July 27 weights-drop breakdown lands in your feed the day it happens.
