Note: This article is adapted from Phala official announcement on the launch of Kimi K3.
Phala just added Kimi K3 to its lineup of models available for confidential inference, and it’s worth paying attention to if you build anything that touches sensitive data. Kimi K3 comes from Moonshot AI and it’s a massive model, 2.8 trillion parameters, built to handle serious coding work, research, and agent tasks that stretch on for a long time. It also understands images natively, not as a bolted on feature, and it can hold up to 1 million tokens of context at once, so it can keep track of an entire codebase or a long stack of documents without losing the thread.
What makes this deployment different from just calling the model through a regular API is where it runs. Phala hosts Kimi K3 inside a GPU trusted execution environment, which is a secure enclave that keeps your prompts, your documents, and everything you send to the model shielded from the infrastructure around it. Nobody operating the servers can see what you’re sending or what comes back, and the whole setup can be independently verified through attestation, so you’re not just taking someone’s word for it.
Why This Matters for Builders
If you’re working with proprietary code, internal research, client data, or anything you wouldn’t want sitting in the clear on a third party server, this changes what’s possible. You get a frontier level model capable of long coding sessions, deep research, and vision based agent work, but without giving up privacy to get there. That combination has been hard to find until now. Most powerful models require you to trust the provider with your data. Here, the trust is replaced with verification.
This is useful for teams building coding agents that need to work across large repositories over long sessions, for researchers combining text and visual data, and for any application where the cost of a data leak is too high to risk. You get the intelligence of a frontier model with a much smaller attack surface.
Getting Started
Kimi K3 is available now on Phala Cloud under the model name moonshotai/kimi-k3, priced at $3 per million input tokens and $15 per million output tokens.
The API documentation is available on Phala’s docs site here: https://phala.com/models/moonshotai/kimi-k3, and if you want to understand the architecture behind the model itself, Moonshot AI has published the full technical report. Check it here: https://arxiv.org/abs/2607.24653
Top comments (0)