Moonshot AI Officially Launches Kimi K3: World's Largest Open-Source Model with 2.8 Trillion Parameters
N.R. Finch
Moonshot AI released its flagship open-source model Kimi K3 on July 16, with 2.8 trillion parameters — the world's largest open-source model; this signals that Chinese AI labs are using open-source plus aggressive pricing to challenge the moat and pricing power of America's closed-source leaders.
2.8 trillion parameters — how big is this model, really?
Kimi K3 uses a MoE architecture — mixture of experts, where the model holds many specialist sub-networks but activates only a fraction per query, saving compute while scaling up. Total parameters: 2.8 trillion. It supports a 1-million-token context window and has native vision.
This means → K3's parameter count already exceeds market estimates for Anthropic's flagship Opus 4.8 (roughly 1.5–2 trillion), opening a gap on sheer model size.
K3 ships as open-weight: developers can download, deploy locally, and modify it for free, with APIs and documentation available from day one. In plain terms = you get the model itself at zero cost and can reshape it however you like.
Does performance actually match up?
Official benchmarks show K3 reaching open-source state of the art on software engineering (SWE-bench Verified), competitive coding (LiveCodeBench), agent tasks (Tau2), and math reasoning (AIME), with some scores approaching or matching top closed-source models.
The Financial Times, citing people familiar with the matter, reports that Moonshot internally positions K3 as a direct rival to Anthropic's Opus 4.8, expecting overall performance to match or exceed it across multiple mainstream benchmarks.
K3 is still not expected to surpass Fable — the frontier model Anthropic paused over safety concerns. This reflects a remaining gap between catching the strongest released closed-source model and reaching the unreleased frontier.
Open-source plus cheap pricing — whose pricing power is under attack?
Anthropic has announced a roughly 50% price hike for Opus 4.8 in September, to $3 per million input tokens and $15 per million output tokens. Moonshot's earlier K2.6 was priced at roughly one-third of Opus 4.8.
This means → K3's open-source play directly compresses the pricing room for closed-source models. Silicon Valley and European companies are already shifting to cheaper Chinese models to cut spending on U.S. lab technology.
In plain terms = if a free or ultra-low-cost model performs well enough, there is no reason to keep paying premium prices for closed-source alternatives. The competitive axis shifts from "who has the strongest model" to "who has the largest developer ecosystem."
Does the "China is eight to twelve months behind" line still hold?
Andreessen Horowitz co-founder Marc Andreessen previously said Zhipu's GLM-5.2 was "the first Chinese AI model to match or exceed a major U.S. lab's public model with no compromises."
DeepSeek, Zhipu, and Moonshot have all launched high-performance open-source models in rapid succession. This reflects a real-time rewrite of the China–U.S. AI gap narrative — the industry consensus of "eight to twelve months behind" now faces reassessment.
The true test, however, is K3's real-world benchmark results and developer-ecosystem traction after launch — that is the checkpoint for whether this "gap is closing" narrative holds up.
How large is the valuation gap still?
Moonshot is raising a new round at an implied valuation of roughly $31.5 billion; DeepSeek is also fundraising at roughly $71 billion.
By contrast, Anthropic's latest valuation is roughly $96.5 billion and OpenAI's roughly $85.2 billion — the valuation gap between top Chinese and U.S. AI labs remains stark.
This means → capital markets have not fully bought the "China has caught up" narrative. Whether the valuation gap narrows depends on how K3 and similar models perform in real-world deployment and ecosystem building — not just benchmark scores.
Content is for reference only, not financial advice.