Talos Linux for AI · Talos Cupar
Your models. Your data.
Your infrastructure.
Serve, fine-tune and ground open models on GPUs you own, without your data leaving the building. Talos Linux for AI is the open-source, immutable operating system underneath. Talos Cupar is the commercial platform that runs it.
Talos Cupar and Talos Linux for AI · General availability March 2027
What you can build
Select. Train. Customize. From an open model to an assistant that knows your business.
Select
Open-weights models
Pick open-weight models from a managed library, pinned to an exact revision and verified by SHA-256. Serve them through an OpenAI-compatible API on NVIDIA or AMD GPUs you already own.
Train
SFT LoRA fine-tunes
Teach the model your language with per-tenant LoRA adapters trained on fenced GPUs. Automatic evaluation splits show before and after quality on your own held-out data.
Customize
RAG over your documents
Ground answers in your own document library with numbered citations, and connect live systems of record through MCP so agents can read and act.
How it works
The scaffolding is generic. What you bring makes it yours.
Talos provides the full path from a model to a working assistant: serving, training, retrieval and connectors. What makes the result yours are the three things only you have, your instructions, your documents and your databases. Read it from the bottom up.
Delivered through Talos Cupar
Chat
Conversational chat with memory, grounded and cited
Agentic AI
Retrieve, reason and act across your systems
Scaffolding provided by Talos
Inference harness
vLLM, OpenAI-compatible API
SFT LoRA
Training pipeline
RAG
Retrieval and vector database
MCP
Connectors and tools
You bring your customization
AI models
Open weights or frontier
Training instructions
Prompts and responses
Enterprise documents
PDF, Word, wikis
Enterprise databases
Via connectors
Declarative
Apply a config, watch it converge
GPUs, storage, model libraries and the revisions to pull are declared as documents. Apply them and the node converges, with no golden images and no shell runbooks.
Reboot-proof
Upgrades stay calm
Models re-verify in place rather than downloading again, and paired nodes roll operating system and driver upgrades one at a time with no platform downtime.
Observable
Status from one place
GPU allocation, model pulls with per-file progress, training phases and retrieval health all surface as first-class status.
Talos model training
Teach the model your voice via SFT LoRA.
01
Your data
Prompt and response pairs, in simple JSONL.
02
Validate
Format checked and an evaluation split made.
03
Train
Adapters are created for the models you selected.
04
See proof
A before and after report with real samples.
05
Go live
Served under its own name.
Talos RAG
Ground it in your documents using retrieval augmented generation.
01
Load
PDF, Word and HTML documents go into your library.
02
Ingest
Documents are parsed, chunked and embedded on GPU.
03
Ask
Hybrid vector search and reranking find the best passages.
04
Answer
The model answers from those passages, with numbered citations.
05
Verify
Every response links back to its source.
Talos MCP
Connect it to your systems using Model Context Protocol.
Agents ground answers in live enterprise data and take approved actions, such as creating a ticket or updating a record. Every call is scoped to a role and logged.
01
Connect
MCP servers for your own systems.
02
Discover tools
Agents see only approved actions.
03
Retrieve and act
Read and write systems of record.
04
Guard
Role-scoped, audited calls.
05
Perform
Workflows across systems.
Talos Cupar
One console for the AI lifecycle.
Talos Cupar is the commercial platform for running AI on private infrastructure. It turns the building blocks above into an operated service, with one place to decide what runs, on which silicon, trained on what, and who can reach it.
Select
Models and inference engines
Choose the models and serving engines that run, each pinned to an exact, verified revision.
Allocate
GPU capacity
Allocate GPUs through sharding and slicing, and repurpose capacity as demand changes.
Configure
Fine-tuning, RAG and MCP
Provision fine-tuning including LoRA, retrieval pipelines and MCP connectors from one interface.
Deliver
Chat and agentic AI
Ship conversational chat with memory and tool-calling agents, with grounded answers and numbered citations. Clients pick any base model or tenant adapter by name.
Control, privacy and compliance
Sovereignty is about control, not location.
Keeping data in-country is not the same as controlling it. What settles sovereignty is who holds the keys, who can reach the machines, where models are trained and which law applies. Talos Linux for AI has a concrete answer for each.
Who holds the keys
Weights stay on your storage
Models, adapters, vectors, documents and prompts live on your hardware, in your jurisdiction. Nothing is escrowed with a provider and nothing phones home.
Who has access
Nothing to log into
No shell, no SSH, no package manager. Talos Linux exposes one mutual-TLS API, with role-based separation between operating nodes and consuming AI.
Where models are trained
On your GPUs, on your data
Fine-tuning runs on your own fenced GPUs. Every adapter carries a full lineage, from base revision and dataset hash to recipe and evaluation scores.
Tenant isolation
Clean walls between customers
Tenants are isolated at the retrieval layer and the adapter layer. One platform serves many customers without mixing their data.
Supply chain
Evidence for the auditor
Models arrive revision-pinned and hash-verified. Talos Enterprise Linux adds SBOMs, curated VEX, signed build attestations, FIPS 140-3 compliant builds, CVE SLAs and IP indemnity.
Which law applies
Runs disconnected
The whole stack runs in your jurisdiction on hardware you own, and runs air-gapped when the rules require it. Nothing meters.
The difference, side by side
A platform, from silicon to chat.
Typical GPU stack
Hand-built and hard to reproduce
- General-purpose OS, hand-configured, drifts over time
- Scripts download models to ad-hoc paths
- One model per GPU budget line
- Training and customization bolted on with a second vendor
- Cloud API where data leaves and costs meter
Talos Linux for AI
One declarative, verifiable platform
- Immutable OS, the whole platform in one declarative file
- Pinned, verified, versioned model library with an API
- Dozens of tenant fine-tunes on one shared base model
- SFT LoRA fine-tuning with before and after evaluation reports
- RAG retrieval, tuning and serving on the same substrate
- MCP role-scoped, audited access to your systems of record
- Your racks, your data, your models, with nothing metered
Availability
Talos Cupar and Talos Linux for AI.
Talos Cupar
The commercial platform for private AI
Full production support and direct engagement with Sidero Labs engineering, from onboarding onward.
General availability · March 2027
Talos Linux for AI
Free and open source, with an enterprise tier
The open-source OS is free under MPL-2.0. Talos Enterprise Linux adds the full AI suite, compliance artifacts and support under one commercial relationship.
General availability · March 2027
See it on your data
An assistant that knows your business.
Talk to us about running Talos Linux for AI and Talos Cupar on infrastructure you own, or get notified when they reach general availability.
Talos Cupar and Talos Linux for AI · General availability March 2027