Talos Linux for AI: private AI on an immutable OS
Sidero Labs has announced Talos Linux for AI, an open-source OS for running private AI on hardware you own, and Talos Cupar, a commercial console that runs on it. Both are planned for general availability in March 2027.
Why we care
Talos Linux is the OS we choose for the clusters we build. It has no shell, no SSH and no package manager, and everything is managed through one API. That matters more when the workload is a model reading private data. Until now, the AI layer on top was ours to assemble. This puts it in the OS.
What is in it
- Model serving: open-weight models pinned to a revision and checked by SHA-256, served on vLLM behind an OpenAI-compatible API. You know exactly which weights are answering.
- Fine-tuning: LoRA adapters per tenant, with an evaluation report before and after.
- Retrieval: search over PDF, Word and HTML, with citations.
- Connectors: MCP connectors scoped by role, with every call audited.
- Upgrades: declarative config. Upgrades re-verify models and roll OS and driver updates one node at a time.
- Isolation: data stays on your storage, tenants are isolated, and it can run air-gapped.
- GPUs: NVIDIA or AMD. Cupar adds splitting GPUs across workloads.
On our privacy levels, this is level 6: your hardware, and no one else sees the prompts.
Editions
- Talos Linux: open source, MPL-2.0.
- Talos Enterprise Linux: adds SBOMs, signed build attestations, FIPS 140-3 builds and CVE response times.
- Talos Cupar: the commercial console, with support.
Also on our list: Talos Director
Sidero also has Talos Director, a commercial platform that runs virtual machines and Kubernetes clusters side by side on Talos, with live migration, snapshots and a tool to convert existing VMs. Most on-prem environments still run VMs next to their clusters, so one platform for both is worth a close look. It is in limited availability now, with general availability planned for January 2027.
Next
We hope to get access to both as soon as possible so we can evaluate them on real hardware. We will post what we find here.
Source: Talos Linux for AI and Talos Director, Sidero Labs.