Eighteen months ago, Moonshot AI was a cautionary tale. DeepSeek's R1 release gutted its user base, dropped it from third to seventh place among China's chat apps, and left the company's future looking shaky.
This week, it released the largest open-weight AI model the world has ever seen, and put its own number on how close that model gets to Anthropic's best.
The number isn't first place. But it's close enough that it's changing the conversation about what "open" AI can do.
So What Actually Is Kimi K3?
Kimi K3 is Moonshot's newest flagship model, launched July 16, 2026, and the company is calling it the world's first open 3T-class model.
A Mixture-of-Experts system with roughly 2.8 trillion total parameters, activating just 16 of 896 experts per task
A 1-million-token context window, aimed at long-horizon coding and agent work
New architecture the company calls Kimi Delta Attention and Attention Residuals, changing how information flows across a sequence and through the model's depth
Ships in two variants at launch: K3 Max for chat and agent tasks, and K3 Swarm Max for large-scale parallel processing
Available now through the Kimi app and Kimi Code, with full open weights scheduled for July 27, 2026
Pricing undercuts most Western competitors: $0.30 per million tokens for cache-hit input, $3.00 for cache-miss, and $15.00 for output, with Moonshot reporting cache-hit rates above 90% on coding workloads.
The Comparison Everyone's Making
Moonshot didn't shy away from naming names. In its release materials, the company said K3 performed "competitively" with Anthropic's Claude Fable 5, currently one of the most capable AI models publicly available, and said it "substantially outperformed" Anthropic's Claude Opus 4.8 and OpenAI's GPT-5.6 Sol and GPT-5.5 on several tasks.
That's a company's own framing, worth reading with the usual grain of salt. But independent evaluators are landing in a similar place. Third-party benchmarking from Artificial Analysis found K3's performance comparable to GPT-5.5 and Claude Opus 4.8, particularly on complex multi-step tasks, while ranking first on web interface building and second overall, behind Fable 5 but ahead of GPT-5.6 Sol.
Moonshot itself is direct about where the ceiling still sits. The company's technical documentation states K3 leads on benchmarks like BrowseComp, SWE Marathon, and OmniDocBench, but trails Fable 5 on FrontierSWE and HLE-Full, two tests built around frontier-level reasoning.
| Date | What Happened | What It Means |
|---|---|---|
| January 2025 | DeepSeek's R1 model disrupts the Chinese AI market. | Moonshot's Kimi app falls from third to seventh place in monthly users. |
| January 2026 | Moonshot closes a $500M Series C at a $4.3B valuation, earmarked for K3. | The company bets its comeback on a single, much larger model. |
| June 9, 2026 | Anthropic releases Claude Fable 5 and Mythos 5. | Sets a new bar for frontier model performance. |
| June 12-30, 2026 | Anthropic suspends Fable 5 and Mythos 5 access under U.S. export controls, then restores it once the controls are lifted. | Highlights that closed, API-only models can be switched off by regulators. |
| July 16, 2026 | Moonshot releases Kimi K3, a 2.8T-parameter open model. | The largest open-weight model yet, positioned near Fable 5's tier. |
| July 27, 2026 (expected) | Full K3 weights are scheduled to post publicly. | Anyone will be able to download and run the model directly. |
Why the Timing Isn't an Accident
K3 arrives about two weeks after the U.S. Commerce Department lifted export controls that had briefly forced Anthropic to pull Fable 5 and Mythos 5 from worldwide access, after the company said it couldn't verify user nationality in real time to comply with the restrictions. Anthropic restored access on July 1 once it added the necessary safeguards.
That episode, however brief, made a point Chinese labs have been arguing for a year: a model that lives behind an API can be switched off by a government decision. A model whose weights are already on your own servers cannot. Analysts have noted that a weights release, once made, isn't something any regulator can reverse after the fact.
Moonshot's release is also timed just ahead of the 2026 World Artificial Intelligence Conference in Shanghai, giving the company a marquee moment to reintroduce itself after a rough stretch.
Kimi K3 and Claude Fable 5: FAQ
Kimi K3 is Moonshot AI's newest flagship model, released July 16, 2026. It's a Mixture-of-Experts model with roughly 2.8 trillion total parameters, a 1-million-token context window, and new Kimi Delta Attention architecture. Moonshot calls it the world's first open 3T-class model.
No. Moonshot's own materials say K3 performs "competitively" with Fable 5 but still trails it overall, along with GPT-5.6 Sol. On several individual benchmarks, K3 reportedly ranks ahead of Claude Opus 4.8 and GPT-5.5, and close behind Fable 5.
K3 launched July 16, 2026 on the Kimi app, Kimi Code, and API. The full open-weight release, letting anyone download and run the model locally, is scheduled for July 27, 2026.
Claude Fable 5 is one of Anthropic's Mythos-tier models, first released in June 2026, and widely regarded as among the most capable AI systems publicly available. Its sibling model, Mythos 5, is restricted to a small number of organizations under Anthropic's Glasswing program due to its cyber-related capabilities.
It means developers can download K3's trained parameters and run or modify it themselves, rather than only accessing it via a paid API like Fable 5 or GPT-5.6 Sol. At 2.8 trillion parameters, Moonshot says K3 will be the largest openly downloadable model once weights post on July 27.
K3 lands about two weeks after the U.S. lifted export controls that had forced Anthropic to briefly suspend Fable 5 and Mythos 5 worldwide. The episode underscored a point Chinese labs have made for a year: an open-weight model, once downloaded, can't be switched off by any government.
Jans Bock-Schroeder
Publisher & Founder of AI Angst
Coming from the world of art, photography, and the luxury market, Jans launched AI Angst in 2025 to explore the cultural, ethical, and psychological impacts of artificial intelligence. His work bridges creative vision with critical technology analysis, offering clarity in an era of rapid technological change.
Sources and Citations
This article is based on the following primary sources, published July 16-17, 2026:
-
Fortune — "Moonshot's Kimi K3 pushes Chinese AI into Fable-level territory" (July 16, 2026)
Primary source for Moonshot's own performance claims against Fable 5, Opus 4.8, and GPT-5.6 Sol.
https://fortune.com/2026/07/16/moonshots-kimi-k3-pushes-chinese-ai-into-fable-level-territory/ -
VentureBeat — "China's Moonshot AI releases Kimi K3, the largest open-source model ever" (July 16, 2026)
Source for Moonshot's history, DeepSeek disruption context, and market comeback narrative.
https://venturebeat.com/technology/chinas-moonshot-ai-releases-kimi-k3-the-largest-open-source-model-ever-rivaling-top-u-s-systems -
MarkTechPost — "Moonshot AI Releases Kimi K3: A 2.8 Trillion Parameter Open MoE Model" (July 16, 2026)
Source for technical architecture details, pricing, and benchmark specifics.
https://www.marktechpost.com/2026/07/16/moonshot-ai-releases-kimi-k3-a-2-8-trillion-parameter-open-moe-model-with-kimi-delta-attention-and-1m-context/ -
CNBC — "China's Moonshot AI unveils Kimi K3 model it says rivals OpenAI, Anthropic" (July 17, 2026)
Source confirming Moonshot's own statement that K3 trails Fable 5 and GPT-5.6 Sol overall.
https://www.cnbc.com/2026/07/17/moonshot-ai-kimi-k3-model-openai-anthropic-china.html -
Technology.org — "Moonshot Unveils Kimi K3, a 2.8 Trillion-Parameter Open-Weight AI Model" (July 17, 2026)
Source for the export-control timing context and analyst commentary on open-weight releases.
https://www.technology.org/2026/07/17/moonshot-kimi-k3-open-weight-ai-model/
Published: July 17, 2026. Sources verified at time of publication. All external links open in a new tab. Benchmark comparisons reflect the releasing companies' own reported figures unless otherwise noted as independent.
