Kimi K3: The Complete Developer Guide

Moonshot AI just put out Kimi K3, and the headline number is hard to ignore: 2.8 trillion parameters, making it the largest open-weight model anyone has shipped. For context, that's not a small step up from the last big open release — it's Moonshot doing what Moonshot has done for a year now, which is push the ceiling on open-model scale in nine of the last twelve months. Together AI is hosting it directly with Moonshot's team, and the pitch is straightforward: this is the first open-weights model built to compete with GPT 5.6 Sol and Claude Fable 5 on frontier coding and reasoning work, not just catch up to last year's closed models.

What's actually new isn't just size, it's how Moonshot kept a model this big trainable at all. K3 activates only 16 of 896 experts per token — roughly 2% — through something called Stable LatentMoE, and that kind of sparsity usually makes training brutally unstable. So they layered in a handful of fixes: Quantile Balancing to route experts without fragile hand-tuned thresholds, a per-head version of the Muon optimizer, and tweaks to activation and attention (SiTU, Gated MLA) that keep things from falling apart at scale. On top of that sits Kimi Delta Attention, a hybrid linear-attention setup that's the reason K3 is the first Kimi model to handle a full 1-million-token context, plus a mechanism called Attention Residuals that lets the model pull specific representations from earlier layers instead of just piling everything on as it goes deeper.

For developers, the more interesting part might be the API design. K3 ships with a reasoning_effort dial — low, high, or max — and you can kill thinking entirely for instant, cheaper responses. It supports OpenAI-compatible tool calling, vision input up to 4K images, structured JSON output, and genuinely automatic context caching if you keep your long prompts byte-identical between calls. There's also a