SAGE
A suite of highly-efficient reasoning models scaling from edge deployment to autonomous enterprise workflows.
Model Variants
Currently in active pre-training. Magnus is our experimental frontier model, developed entirely from the ground up to explore the absolute limits of sparse MoE logic and multi-modal synthesis. It is explicitly designed for domains where rigorous correctness and deep ideation outweigh immediate latency. Early checkpoints are currently undergoing extensive internal safety evaluation.
Explore SAGE Magnus →The autonomous workflow powerhouse. Actus brings a new level of consistency to multi-step agentic planning and professional software engineering.
Explore SAGE Actus →The lightweight, extremely fast reasoning models. Designed for rapid inference on standard consumer hardware without sacrificing analytical depth. Perfect for real-time applications and localized parsing.
Explore SAGE Celer →Announcements
SAGE Magnus has officially completed its Phase 1 pre-training. We clarify its MoE architecture, our open-source data strategy, and how we secured compute for a from-scratch training run.
Read more →Magnus is currently reaching the end of its Phase 1 pre-training. We are preparing to conduct extensive internal safety testing and ethical red-teaming to ensure our upcoming reasoning system is built on a foundation of absolute transparency and human alignment.
Read more →Announcing the beginning of development for SAGE Magnus, our upcoming reasoning engine. Designed to eventually feature an unbound computation budget, Magnus will allow researchers to allocate immense "Thinking time" for unprecedented problem-solving density.
Read more →The evolution of SAGE-32B into a fully realized agentic reasoning platform. Building on the intelligence of its predecessor, it brings new levels of reliability and precision to coding, agents, and enterprise workflows.
Read more →A 32 billion parameter language model that focuses on agentic reasoning and long range planning tasks. SAGE-32B achieves higher success rates in multi-tool usage scenarios compared to similarly sized baseline models.
Read more →Announcing the first reasoning models in the SAGE series, combining extreme efficiency with advanced problem-solving capabilities. Featuring SAGE Celer Low 2.5 (3B), SAGE Celer Mid 2.5 (8B), and SAGE Celer High 2.5 (14B).
Read more →An early internal iteration of our IDA-based reasoning models, paving the way for advanced efficiency. While less advanced than our subsequent public releases, version 1.5 demonstrated foundational capabilities that validated our approach to extreme efficiency without sacrificing meaningful problem-solving abilities.
Read more →Built For Deep Reasoning
SAGE models represent a fundamental shift in how we approach reasoning parameters inside language models. By explicitly modeling logic as a first-class function inside the architecture, the SAGE family outperforms dramatically larger proprietary tools natively.
Inverse Reasoning
Our models are capable of dissecting their own logic paths backwards during generation to self-correct hallucinations before outputting to the user.
Local-First Capability
By optimizing parameter logic rather than simply relying on scale, you can run state-of-the-art intelligent processing completely offline through your own GPU constraints.
Performance Highlights
On agentic reasoning benchmarks including MMLU-Pro, AgentBench, and MATH-500, SAGE-32B achieves higher success rates in multi-tool usage scenarios compared to similarly sized baseline models.
| SAGE-32B | Qwen2.5-32B | Llama-3 (70B) | GPT-4 Turbo | |
|---|---|---|---|---|
General understanding MMLU-Pro | 79.3% | 71.5% | 68.9% | 63.7% |
Math & logic MATH-500 | 91.8% | 78.9% | 68.0% | 72.6% |
Agentic tracking AgentBench | 73.1% | 58.4% | 62.1% | 85.0% |
Graduate reasoning GPQA | 48.0% | 50.5% | 51.0% | 53.6% |
Instruction following IFEval | 84.5% | 81.2% | 78.5% | 86.0% |
Source: SAGE-32B: Agentic Reasoning via Iterative Distillation
ACUMEN Benchmark
SAGEA InternalACUMEN is our composite evaluation framework measuring models across three axes: intelligence (ACUMEN-I), agentic capability (ACUMEN-A), and inference efficiency (ACUMEN-E). Unlike standard benchmarks, ACUMEN rewards efficiency as a first-class dimension — SAGE models consistently lead in ACUMEN-E due to their highly optimized inference footprint.
Read the methodology →Source: ACUMEN Leaderboard · March 2026
| Model | ACUMEN-I | ACUMEN-A | ACUMEN-E | Composite |
|---|---|---|---|---|
| SAGE Actus 2.4 | 80.1 | 66.1 | 85.6 | 71.8 |
| SAGE Celer High 2.5 | 66.1 | 63.4 | 88.9 | 70.7 |
| SAGE Celer Mid 2.5 | 59.4 | 57.1 | 92.1 | 66.6 |
| Claude Sonnet 3.5 | 81.2 | 83.1 | 52.1 | 74.6 |
| GPT-4 Turbo | 77.9 | 81.9 | 47.8 | 72.1 |
ACUMEN-E scores reflect Quality-per-FLOP on standardized hardware. Higher is better across all columns.
Availability
For individual developers and enthusiasts who want to experiment with local reasoning models, the SAGE Celer family is available to run offline via Ollama and standard GGUF executors.
For research laboratories and select enterprise partners requiring absolute state-of-the-art reasoning, SAGE Magnus early access will be granted on a case-by-case basis through our research partnership program as development continues.
For business users and those building complex agentic solutions, SAGE Actus is available on demand for enterprise customers only. Access requires contacting our team directly. The infinite reasoning mode is included as part of an enterprise agreement.
Contact us to discuss enterprise access to SAGE Actus.
Use cases
The SAGE family covers every base, from absolute edge computing logic and zero-latency interfaces to fully orchestrating autonomous AI agent chains designed for the highest stakes workflows.
On-device processing & Real-time edge swarms
SAGE Celer runs entirely on your local machine, ensuring sensitive personal data and proprietary source code never traverse a network. With heavily optimized first-token latency, it natively handles voice-to-voice interfaces and robotics routing with ease.
Frontier Research & Complex Discovery
The planned architecture for Magnus is designed for domains where correctness, deeply novel ideation, and multi-step theorem proving outweigh immediate latency. It will iteratively verify its logic and formulate sub-problems to ensure absolute precision.
Advanced coding & Autonomous workflows
SAGE Actus confidently delivers production-ready code with minimal oversight. It handles longer, more complex task chains with fewer errors, respecting JSON schemas out-of-the-box and dynamically adapting its pipeline as constraints organically emerge.