Cosine

Most coding assistants are built on general-purpose models that happen to be good at code, and run in someone else's cloud. For a startup that is fine; for a bank, a defence contractor or a critical-infrastructure operator, neither assumption holds. Cosine is built for exactly those organisations.

It describes itself as a sovereign AI lab, training specialist models, its Lumen family, purpose-built to write software rather than to also compose poetry, and designed to run wherever your compliance requires, up to a fully air-gapped deployment inside your own perimeter. It also reaches enterprise and niche languages mainstream tools neglect.

This review covers what makes the Lumen models different, how deployment scales from cloud to air-gapped, how Cosine approaches pricing, and who it suits against mainstream coding tools.

Cosine In Four Points For Secure Teams

• Cosine is a sovereign AI lab that trains code-specialist models and agents for organisations needing capability inside secure environments.
• Its thesis is specialists over generalists: its Lumen models do one thing, write exceptional software, rather than everything.
• Deployment goes to air-gapped: public cloud, a managed single tenant, or fully inside your own security perimeter.
• It reaches niche languages like COBOL, Fortran, Verilog and complex SQL that mainstream tools handle poorly.

A Sovereign AI Lab With Code-Specialist Models

Cosine's identity is unusual for a coding tool: it is a lab that trains its own models rather than a wrapper over someone else's. That matters because its whole pitch rests on control over the model and where it runs.

The sovereign framing, including a published open letter on sovereign AI for the UK backed by critical-infrastructure executives, signals the audience: organisations for whom sending code to a third-party API is a non-starter, and who need frontier capability they can run themselves.

That is a different proposition from a mainstream editor like Cursor or an assistant like Codeium, which are excellent for most teams but assume a standard cloud relationship Cosine's customers cannot accept.

For a startup, a cloud coding assistant on a general model is fine. For a bank or a defence contractor, neither assumption holds.

What Makes Cosine's Lumen Models Different?

The technical argument is a focused one, and Cosine states it plainly: a model trained exclusively on real production code will beat a generalist at the task developers actually need.

Specialists Over Generalists For Code

Lumen is a family rather than one model, Scout for cheap, on-device work around the coding loop, Outpost for everyday production tasks, and a frontier Sovereign tier for the hardest problems. Every training and evaluation decision is grounded in one objective: better, more maintainable code, not bloated generation.

This specialist stance is the opposite bet from a general code reviewer like Sourcery sitting on top of broad models; Cosine argues the model itself should be a code specialist, which is a coherent thesis worth testing against your own codebase.

Niche Languages Like COBOL And Verilog

Cosine post-trains for enterprise and niche languages, COBOL, Fortran, Verilog, Rust and complex SQL, that mainstream assistants handle poorly. For organisations maintaining decades-old systems, that coverage is not a nice-to-have; it is often the whole reason to look.

It is also designed to reason across complex architectures, legacy infrastructure and long-lived codebases, which is exactly where general tools, tuned on modern open-source code, tend to struggle most.

From Public Cloud To Fully Air-Gapped Deployment

The deployment ladder is central to Cosine's value. You can start quickly in Cosine Cloud with standard controls, move to a managed single-tenant environment for enterprise needs, or run fully air-gapped entirely inside your own security perimeter.

Air-gapped operation is the feature that opens doors mainstream tools cannot. For defence, government, finance and critical infrastructure, the ability to get frontier coding help without code ever leaving the building is the difference between adopting AI and banning it.

You work through a terminal-native CLI with a local-to-remote execution model, external tool access over MCP and multi-agent orchestration, or in the cloud for parallel tasks and collaboration. The trade-off is that this is enterprise infrastructure, not a one-click consumer app.

How Cosine Approaches Pricing And Deployment Tiers

Cosine's pricing follows its deployment model: a managed cloud tier to start, and quoted arrangements for single-tenant and air-gapped deployments where infrastructure and compliance drive the cost.

AInexfinder does not publish plan prices, and sovereign or air-gapped deployments are bespoke by nature. The real cost question is total: not just licences but the deployment, integration and security work that running models inside your own perimeter entails.

For the organisations Cosine targets, that cost is weighed against the alternative of not using AI at all, or using it in a way their compliance forbids, which is why the sovereign option can be worth a premium a general tool never needs to charge.

Evaluating Cosine In A Secure Environment

Because Cosine's value is specialist models plus secure deployment, evaluate it on your own code and your own constraints, not a generic benchmark.

1. Define your real constraint first: is it air-gapping, a niche language, legacy architecture, or all three?
2. Try the Lumen models on a representative task in your actual codebase, including any niche language you depend on.
3. Judge the output on maintainability and correctness, not just whether it compiles.
4. Map the deployment tier, cloud, single-tenant or air-gapped, to what your compliance actually permits.
5. Pilot the CLI workflow with a small team and test the MCP and orchestration against your tools.
6. Pass/fail: Cosine earns adoption only if it delivers frontier coding help within constraints a mainstream tool cannot meet.
7. Confirm the security and deployment commitments in writing before committing a regulated workload.

Cosine And Sovereign-AI Terms

Sovereign AI

AI an organisation can run and control within its own jurisdiction and infrastructure, rather than depending on a third-party cloud.

Air-gapped

A deployment that runs entirely inside your own security perimeter, with no connection to outside networks.

Lumen

Cosine's family of code-specialist models, spanning the Scout, Outpost and Sovereign tiers.

Post-trained

Further training a base model on specific data, here real production code and niche languages, to specialise it for software.

Cosine Sovereign-AI Facts