We're introducing SAGE, a family of reasoning models built on months of research into efficient, interpretable thinking.

Traditional large language models excel at pattern recognition and fluent text generation, but they struggle with tasks that require step-by-step reasoning. SAGE was designed from the ground up to reason clearly and reliably—thinking through problems with transparency and precision.

Reasoning that works

SAGE combines two key insights from our research. First, we developed inverse reasoning—a novel technique that enables models to explain not just their answers, but their thinking process. Rather than generating reasoning forward, inverse reasoning allows SAGE to decompose its own thought chains post-hoc, showing you why specific conclusions were reached over others.

Second, we focused on efficiency. While other labs scale models to hundreds of billions of parameters, we discovered that with the right training approach, smaller models can reason just as effectively. Our 3B, 8B, and 14B variants deliver reasoning capabilities that previously required models 10x their size—without sacrificing speed or hardware accessibility.

What SAGE can do

In our testing, SAGE excels at tasks that demand clear thinking:

• Complex problem-solving — breaking down multi-step logic puzzles, debugging code, and working through analytical challenges.
• Technical writing — explaining concepts clearly, writing documentation, and crafting precise technical responses.
• Code reasoning — understanding codebases, suggesting fixes, and explaining architectural decisions.
• Research and synthesis — connecting ideas across domains, summarizing complex material, and identifying patterns.

Made for your hardware

Unlike models that require cloud infrastructure, SAGE runs locally on consumer hardware. Our 3B model runs smoothly on a modern laptop. The 8B variant works well on a single GPU. Even our largest 14B model fits on standard hardware without quantization. This means your reasoning stays on your machine—private, fast, and fully under your control.

Months of research, one family of models

SAGE represents the culmination of extensive work on reasoning architectures, training techniques, and interpretability. We published our findings on inverse reasoning, developed novel training approaches for efficiency, and spent months testing with early partners to refine the models before this release.

We believe reasoning is fundamental to useful AI. Whether you're building tools for your team, integrating reasoning into applications, or exploring what AI can do, SAGE is designed to be reliable, transparent, and accessible.

The models are available today, ready to run on your hardware and integrate into your workflows.