Kimi K3: What Moonshot AI’s Giant Open-Weight Model Changes

The largest open-weight model anyone can actually download

Moonshot AI’s Kimi K3 landed in July 2026 and immediately became the biggest open-weight model on the leaderboard – a 2.8-trillion-parameter Mixture-of-Experts model that anyone can download, self-host, or fine-tune under an open licence. The headline number is huge, but the architecture is the more interesting part: Kimi K3 activates just 16 of its 896 experts per token, so the compute cost per request looks nothing like the parameter count suggests.

Why 2.8 trillion parameters doesn’t mean 2.8 trillion parameters of compute

Mixture-of-Experts models split their weights into many small specialist sub-networks (“experts”) and a router that picks a handful to run for any given token. Kimi K3’s 896 experts, with only 16 active at once, means the effective compute per token is much closer to a model a fraction of its total size – which is exactly why MoE has become the default architecture for anyone trying to push open-weight quality up without making inference unaffordable.

• Total parameters: 2.8 trillion, spread across 896 experts
• Active parameters per token: only 16 experts are routed to for any given token
• Context window: 1 million tokens – enough for entire codebases or long document sets in a single prompt

Where it sits next to Qwen3.8, GLM-5.3 and the rest

Kimi K3 arrives into a crowded open-weight field. Qwen3.8 Max still leads the September 2026 benchmark leaderboard by raw score, and GLM-5.3-Flash recently became the first natively multimodal GLM-5 release. Kimi K3’s pitch isn’t topping every benchmark – it’s sheer scale and context length, which matters most for workloads that genuinely need to reason over huge amounts of text in one pass, like whole-repository code review or long-document analysis, rather than short chat-style prompts.

Can you actually run it yourself?

Not on a single consumer GPU, and probably not on a homelab box either. A 2.8-trillion-parameter MoE model – even with only 16 experts active per token – still needs the full weight set resident somewhere, which puts it firmly in multi-GPU server territory rather than a laptop or a Raspberry Pi. For most self-hosters, Kimi K3 is more interesting as an API-hosted option (several inference providers added it within weeks of release) than as something you pull with ollama run.

If you do want to experiment locally, the practical path is a heavily quantised build through a runtime that supports MoE offloading, accepting a real quality trade-off in exchange for fitting it on hardware you actually own.

What this means if you’re choosing a model

• Need the largest possible context window for whole-codebase or long-document work, and don’t mind an API bill: Kimi K3 is worth evaluating
• Want the best benchmark score for general reasoning and coding: Qwen3.8 Max remains the one to beat as of September 2026
• Actually self-hosting on hardware you own: smaller, purpose-sized models will still serve you better day to day – Kimi K3’s scale is solving a different problem

The bigger trend worth watching

Six new models shipped in the first week of September 2026 alone, and Chinese labs – Moonshot, Alibaba (Qwen), Z.ai (GLM) – now account for the majority of tokens routed through services like OpenRouter. Kimi K3 is one data point in a genuinely fast-moving open-weight landscape: the gap between what you can self-host and what the biggest labs ship keeps narrowing, even as the biggest releases themselves keep getting harder to run on anything but a data-centre GPU cluster.