Chinese Moonshot AI Drops Kimi K3, a Mind-Blowing 2.8-Trillion Parameter Giant
A Chinese startup just threw a massive wrench into the open-source AI race, dropping a gargantuan model that makes Western tech giants look like they are hoarding their toys. Let’s see if this beast actually lives up to the hype or just melts servers.
The developers at Moonshot AI decided that size actually matters, packed 2.8 trillion parameters into their new Kimi K3 model, and promised to release its weights to the public before July 27. This monstrous open-weight brain handles text, images, and video while juggling a context window of up to 1 million tokens, which is basically enough memory to digest a whole fantasy novel series in one gulp.
To keep this digital leviathan from tripping over its own feet, the creators introduced a mode called K3 Swarm, which allows multiple AI agents to work in parallel on complex coding and multi-step tasks. In-house benchmarks claim the model already outruns Claude Opus 4.8 in several tests, though it still eats dust behind closed-source heavyweight champions like Fable.
For those who prefer renting brains over hosting them, the API costs sit at $3 per million input tokens and $15 per million output tokens, with an automatic 90% discount applied to cached inputs. The model is currently live on the Kimi web chat, their API platform, and the Kimi Code helper tool.
While Western tech giants keep locking their best models behind corporate paywalls and safety committees, open-weight behemoths from Asia are quietly democratizing insane computational power. The real test will be whether developers can actually run a 2.8-trillion parameter monster on anything short of a small nuclear power plant, or if this open-source gift is just a giant, unusable trophy on GitHub.
Source: Moonshot AI
Comments
This is where the magic happens: AI reads your discussion and rewrites the article based on the most interesting comments. Each strong comment adds points to the meter below. Once the meter is full, the article updates live — no page reload needed.