Models for every depth of engineering work.
Sigilix is an AI lab. We tune our own model line — Boreas, Pyroeis, and Astraeus — memory-native models built for engineering work and tuned to honor the memory and connected integrations of each developer and each organization. The difference is not just what a model knows in general. It is that ours are tuned to respect your memory, your tools, and how your engineering system works — and to keep improving at it over time, as that context compounds.
Light, base, premium, and frontier tiers share one thesis: every tool we build — the repo graph, review history, Slack decisions, Linear context, and CLI sessions — exists to fuel these models with real, verified engineering context, so they get better at your system without forcing you to repeat yourself every time.
Boreas
Fast code work that still understands the repo around it.
Boreas is our light-tier model for quick review, CLI help, and small code changes where speed matters but generic answers are not enough. We tune it around codebase context and memory standards, so everyday interactions across your tools fuel it with how your work is shaped instead of treating each request like a blank prompt.
Best for fast review loops, local CLI questions, routine bug triage, and small edits that need codebase grounding.
Uses memory to preserve recurring patterns, naming decisions, common failure modes, and the habits your team keeps reinforcing.
The Boreas page covers its flash-tier comparisons, effort points, and memory-driven coding benchmark framing.
Pyroeis
Our base model for everyday engineering work.
Pyroeis is the default model for most product teams: grounded enough for multi-file work, fast enough for daily use, and broad enough for the small tasks that keep engineering moving. It carries repository context, user memory, and workflow history into the answer so you can ask for help without restating how everything fits together.
Best for everyday implementation, review follow-up, small refactors, product questions, and work that spans a few files.
Designed for users who need a steady model more often than a frontier one: reliable, practical, and memory-aware.
Its section links directly into the Pyroeis release notes, benchmark framing, and deeper model context.
Astraeus
Repository-scale reasoning for higher-risk work.
Astraeus is our premium model for the work where the cost of missing context is highest: security changes, architecture shifts, complex repairs, and long-running investigations. It keeps PR history, issue trails, CLI sessions, and team decisions available as memory so the model can reason with the system you actually have.
Best for hard bugs, cross-service changes, security-sensitive reviews, and repair loops where the answer depends on history.
The goal is high-effort reasoning that can see the context behind the task, not a detached score chase.
The Astraeus page covers benchmark framing, token-curve context, and the model release story.
Phanes
The private path for guardrails and memory integrity.
Phanes stays private while the release criteria mature around disclosure, provenance, data integrity, and guardrails. The hard part is not only making a stronger memory-native model. It is making sure the model explains when memory shaped an answer, keeps user-visible evidence honest, and never hides context simply because it can.
Best understood as a guardrail and release-readiness case study, not a public model tier you should route normal work through today.
Focus areas include memory disclosure, data boundaries, provenance, and the difference between helpful context and context that should be surfaced.
The Phanes announcement explains why guardrail and data-integrity work happens before a broader release.
One model line. One memory layer.
Sigilix tunes its own models, and builds the review surfaces, CLI workflows, and memory-aware tools that fuel them with the same engineering context. You choose the reasoning depth for the job; the platform keeps the context consistent.
That is the difference between asking a model to guess and asking a system to remember how your codebase works.
The four are capability tiers, not task bots: Boreas, Pyroeis, and Astraeus (with Phanes coming) power the CLI and agent surfaces. Sigilix's PR review is a separate lane — an ensemble of role-specialist models for logic, security, performance, and tests, each with cross-provider fallback.