SAGE Celer is a new series of hybrid models designed to bridge the gap between lightweight deployment and heavy reasoning tasks. Available in highly optimized 3B, 8B, and 14B parameter sizes, Celer redefines what is possible on consumer-grade hardware.

Highlights.

1. Hybrid Thinking Mode: These models feature a unique "Thinking" mode that allows them to dynamically allocate more compute time to complex queries before streaming a response.
2. On-device processing: Celer runs entirely on your local machine or edge server, ensuring that sensitive personal data, financial records, and proprietary source code never traverse an external network.
3. Scalable deployments: Ranging from 3B to 14B parameters, allowing organizations to meticulously match their specific hardware constraints and power limitations with the appropriate reasoning tier.

Model Capabilities.

This release marks a significant step forward in making advanced tool-calling and mathematical reasoning truly ubiquitous. SAGE 2.5 Celer establishes a new state-of-the-art benchmark across a wide range of coding and reasoning capabilities for models within the sub-15B parameter weight class.

The true innovation lies in its hybrid architecture, which allows the model to dynamically scale its compute depending on the cognitive demand of the prompt. During simple conversational turns or basic data extraction, Celer operates with blistering speed, outputting over 100 tokens per second on standard hardware. However, when faced with a complex logic puzzle or a deeply nested coding bug, its "Thinking" mode activates. In this mode, Celer allocates internal, hidden token generation to properly map out a solution, test hypotheses, and verify logic before streaming the final, correct answer to the user.

Additionally, the SAGE 2.5 Celer family has been specifically fine-tuned for structural compliance. It guarantees valid JSON output when constrained by schemas, making it exceptionally reliable as an invisible cognitive backend in traditional software pipelines. It is highly resistant to quantization degradation, maintaining over 95% of its reasoning capabilities even when compressed to 4-bit (GGUF/AWQ) formats.

Use Cases.

• On-Device Copilots: Deploy the 3B variant natively onto edge hardware or within Electron applications to provide robust, offline code completion and contextual chat without invoking expensive cloud endpoints.
• Low-Latency RAG Pipelines: The 8B model is uniquely positioned to handle massive Retrieval-Augmented Generation workflows. It can digest dozens of retrieved documents in milliseconds, outputting cited, factual responses instantly for customer support automation.
• Edge-Based Data Routing: Utilize Celer's perfect structured JSON output to analyze unstructured incoming data streams (like emails or user feedback) and intelligently route them to the correct department databases with zero human intervention.