On July 16, 2026, Moonshot AI released Kimi K3, a 2.8 trillion parameter model the company says is the largest open-weight model ever put into the world, roughly 75% larger than DeepSeek V4 Pro, the previous largest widely used open model (Bloomberg, VentureBeat). Kimi K3 is Moonshot’s clearest bid yet to match frontier capability while giving the weights away.
What Is Kimi K3?
Kimi K3 launched via API first, with full weights scheduled for release under a Modified MIT license by July 27, 2026, about ten days later (Fortune). It ships with a native 1,048,576-token (roughly 1 million token) context window and multimodal input, putting it in the same context-length class as the largest Western models.
How Does Kimi K3 Perform Against GPT-5.6 and Claude Fable 5?
K3 beats Claude Fable 5 on the Frontend Code Arena coding benchmark, and overall it trails only GPT-5.6 and Claude Fable 5 among frontier models (Tom’s Hardware, CNBC). Moonshot prices API access at $3 per million input tokens and $15 per million output tokens, with cache-hit input dropping to $0.30 per million (BenchLM.ai), a small fraction of what comparable closed frontier models typically charge.
Why Moonshot Keeps Giving Away Frontier-Class Weights
Moonshot has raised roughly $1.5 billion total, with its valuation climbing from $2.5 billion to $4.3 billion by early 2026 (VentureBeat). Its predecessor, Kimi K2, released in July 2025, built its reputation on strong coding performance at a lower price than Western rivals. K3 continues that same price-and-openness strategy, just at a far greater scale.
The strategic logic is straightforward even if it looks unusual from a Western licensing mindset: giving away frontier-class weights builds developer adoption and ecosystem lock-in that a closed model can’t buy as quickly, while hosted API access still generates direct revenue from anyone who’d rather not run a 2.8 trillion parameter model themselves.
What This Signals
A near-frontier model, open-weight, priced at a fraction of closed competitors, is a real test of Western labs’ pricing power. It’s also the exact backdrop for two other questions covered elsewhere in this series: whether the US could realistically restrict foreign open-weight models, and how distillation lets smaller teams compress a model like K3’s capability into something they can run themselves.
Frequently Asked Questions
Is Kimi K3 really open source? The API launched first on July 16, 2026, with full model weights scheduled for release under a Modified MIT license by July 27, 2026. That’s a permissive open license, though it isn’t necessarily the formal Open Source Initiative definition, since not all training data or code is published alongside the weights.
How big is Kimi K3 compared to other open models? At 2.8 trillion parameters, Kimi K3 is described by Moonshot AI as the largest open-weight model released to date, roughly 75% larger than DeepSeek V4 Pro, the prior largest widely used open model.
How does Kimi K3 compare to GPT-5.6 and Claude Fable 5? Kimi K3 trails only GPT-5.6 and Claude Fable 5 among frontier models on overall benchmarks, and it actually beats Claude Fable 5 on the Frontend Code Arena coding benchmark.
What does Kimi K3 cost to use through Moonshot’s API? Moonshot prices Kimi K3 at $3 per million input tokens and $15 per million output tokens, with cached input tokens dropping to $0.30 per million, a fraction of typical closed frontier model pricing.
Who is behind Kimi K3? Kimi K3 comes from Moonshot AI, a Beijing-based AI startup that has raised roughly $1.5 billion and saw its valuation climb from $2.5 billion to $4.3 billion in early 2026.
