Kimi K3 Ships Open Weights at 2.8 Trillion Parameters: What It Means for Business AI
3 August 2026
Moonshot AI published open weights for Kimi K3 on 27 July, making it the largest freely available model released to date at 2.8 trillion parameters. It handles text, images, and video in a single model, supports a 1 million token context window, and ships in MXFP4 quantisation at roughly 1.4 TB. Moonshot's benchmarks place it just behind the current frontier models from Anthropic and OpenAI, while claiming it is two to three times cheaper to run.
[Source: Tom's Hardware]
Why This Matters
The gap between open and closed models keeps narrowing. A year ago, choosing an open-weight model meant accepting a visible quality drop. Kimi K3 trailing the frontier by a small margin changes the calculation for any workload where "very good" is enough, which is most business workloads.
Self-hosting is now a real option, but not a cheap one. Moonshot cites eight Nvidia H100 GPUs or more than 2 TB of VRAM to run it yourself. That is an infrastructure commitment most SMEs will not make directly. The practical route is a hosted provider serving the open weights, which gives you competitive pricing and the ability to switch providers without changing models.
Portability is the real prize. Open weights mean the model cannot be deprecated, repriced, or restricted underneath you. For European businesses with GDPR obligations or data residency requirements, an open model you can run in an EU region, or on your own hardware if it ever matters, removes a category of risk that closed APIs cannot.
Our Take
The headline number is not the interesting part. Very few businesses need a 2.8 trillion parameter model, and almost none should be self-hosting one. What matters is what a release like this does to the market: it puts steady downward pressure on frontier pricing and it makes open weights a credible default rather than a compromise.
Our advice to clients has not changed much. Match the model to the task, keep your prompts and evaluation harness portable, and avoid architectures that assume one specific provider. When a model this capable becomes freely available, the businesses that benefit fastest are the ones who can swap it in behind an abstraction layer and measure the difference. The ones locked into a single vendor's tooling will still be reading the announcement six months from now.
For most workloads, the winning setup remains a mid-sized model grounded in your own data rather than the biggest model available. Open weights simply widen the field of models you can choose from, and lower the cost of changing your mind later.
Related reading:
