SAGE OSS represents our core commitment to the open-source community. By releasing high-quality base models and supervised fine-tuned variants, we empower researchers globally to build upon our reasoning breakthroughs directly without arbitrary constraints or restrictive licensing.

Highlights.

1. Truly Open Weights: Accessible architecture allowing researchers full freedom to adapt, study, and fine-tune the model for niche industrial, academic, and enterprise applications.
2. Highly Capable Base Model: Pre-trained on a massive, highly curated multilingual corpus, SAGE OSS demonstrates exceptionally strong performance in English prose, code generation, and complex math.
3. Uncensored Research Base: Stripped of overly rigid system guardrails to ensure utility across medical, legal, and academic domains where precise data extraction is required over restrictive consumer safety filters.

Model Capabilities.

The AI landscape thrives on open collaboration. With SAGE OSS, we are releasing a 12-billion parameter model designed to comfortably sit within the VRAM of a standard consumer GPU (requiring only 24GB of VRAM in fp16, and as little as 8GB quantized) while punching far above its weight class. It sets a robust baseline for instruction adherence, logic, and general knowledge, making it the ideal foundation for any custom deployment or internal enterprise tool.

Rather than relying heavily on parameter bloat or inefficient Mixture-of-Experts routing overhead, SAGE OSS utilizes a pure, dense architecture combined with intensive training data curation. We employed a novel curriculum learning strategy during its pre-training phase. By ensuring that the model develops deep representations of logic, programming syntaxes, and mathematical proofs before being subjected to broad general knowledge, SAGE OSS naturally structures its reasoning. This results in a highly coherent model that follows strict formatting and complex system prompts exceptionally well.

In our rigorous internal evaluations, SAGE OSS consistently outperforms legacy open-source models and even directly competes with highly optimized proprietary models in the 7B to 14B parameter range. It solidifies its place as a reliable, versatile workhorse for open-source intelligence.

Use Cases.

• Academic & Institutional Research: Free from the "black box" nature of API-based models, academic researchers can inspect SAGE OSS's layer activations, experiment with custom attention mechanisms, and conduct foundational AI safety research.
• Specialized Fine-Tuning: Enterprises can leverage Parameter-Efficient Fine-Tuning (PEFT) methods like LoRA to train SAGE OSS on highly proprietary industrial protocols, localized legal frameworks, or specialized coding languages without sharing their data externally.
• Medical & Legal Data Extraction: Because SAGE OSS does not suffer from aggressive consumer-level refusals, it is perfectly suited for parsing sensitive medical transcripts, synthesizing legal contracts, and executing objective data processing in highly regulated environments.