MODELS
Intelligence takes
more than one form.
Speed when the moment matters. Depth when the answer matters. Expertise at scale. Precision in software. Meet the four minds of the Inertial Logic model family.
Versions, architectures, and numbers below are internal research specifications. Speed figures are design targets, not benchmark results. These models are not available for public use.
LYRA / RESPONSIVE INTELLIGENCE
Lyra2.4
The fastest path from signal to understanding.
RESEARCH PROFILE / 2.4Intuition at the
speed of intent.
A small footprint. An unusually long reach. Lyra is the most responsive model of the Inertial Logic model family: interpreting signals, resolving routine questions, and keeping the world moving while deeper reasoning unfolds elsewhere.
- Dense parameters
- 24B
- Context tokens
- 500K
- Target tokens / sec
- 750
- Target first token
- <120 ms
PulseWeight Execution Cascade concentrates attention on the evidence most likely to change an answer. Routine work stays blazing fast, while consequential uncertainty moves to Titan in parallel, bound to Lyra by a linked confidence envelope. The design target is instantaneous responsiveness with parallel execution of deep reasoning.
TITAN / FLAGSHIP REASONING
Titan3.1
Depth is an architecture, not a longer answer.
RESEARCH PROFILE / 3.1Stay with the
hard problem.
Where Lyra answers at the speed of intent, Titan keeps working. It follows competing explanations, carries unresolved questions forward, and turns a large body of evidence into a decision that can be examined, running alongside Lyra so that speed never has to wait for depth.
- Dense parameters
- 420B
- Context tokens
- 1M
- Reasoning branches
- 64
- Deliberation passes
- 2
Gravitas allocates deliberation by uncertainty and consequence. Competing explanations are retained long enough to challenge the leading answer. Through the linked confidence envelope, Titan can change direction mid-turn if Lyra’s confidence starts to drop.
AETHER / SPARSE EXPERT INTELLIGENCE
Aether1.0
A vast field of expertise. A selective path through it.
RESEARCH PROFILE / 1.0A trillion weights.
A precise path.
Aether explores a wider space of expertise without activating the entire model for every token. Its 1.1-trillion-parameter architecture brings specialist reasoning into a common representational frame.
- Total parameters
- 1.1T
- Active / token target
- 88B
- Context tokens
- 2M
- Experts
- 512
Prism Routing pairs the problem with complementary expert groups. A shared residual representation carries context across those selections. The design target is specialist depth without fragmenting the overall reasoning.
ARGO / FORGE CODING INTELLIGENCE
Argo2.2
Every connection has a contract.
RESEARCH PROFILE / 2.2Read the system.
Rewrite the possible.
Argo is the code-specialist profile and terminal-native identity within Atlas Forge. It moves from repository-scale understanding to precise implementation, with the original objective attached to every proposed change.
- Dense parameters
- 96B
- Repository context
- 1M
- Language families
- 32
- Verification gates
- 4
ProofWeight favors proposals supported by executable checks, preserved contracts, and independent review. New evidence can change the preferred implementation. The design target is a result that survives inspection, not just a convincing explanation.
Different scales.
Shared judgment.
A model family is a system of handoffs. Lyra handles immediacy. Titan earns difficult conclusions. Aether opens a broader expert space. Argo turns software intent into an inspectable artifact.