CMSManhattan/JiRackUltra_32b

TEXT GENERATIONPricing:Input $2.72 / Output $4.8Concurrent Unit Cost:2Model Size:32.8BQuant:FP8Context Size:32kTool Calling:SupportedPublished:Aug 3, 2026License:mitArchitecture:Transformer0.0K Open Weights Featherless Exclusive Cold

CMSManhattan/JiRackUltra_32b is a 32.8 billion parameter model built on a DeepSeek R1-32B architecture, refactored with native BitNet features for efficient CPU inference. It utilizes an updated tokenizer with specialized tags for Routing, Media, Vision, Sound, Tool calls, and Robotics, making it highly optimized for domain-specific expert applications. This model excels in scenarios requiring advanced tool calling, robotics integration, and efficient cloud-ready deployments, particularly as an expert model in RAG systems.

Loading preview...

JiRack Ultra 32B: CPU-Optimized Model with Advanced Tooling

CMSManhattan/JiRackUltra_32b is a 32.8 billion parameter model engineered for fast and efficient CPU inference. It is built upon a DeepSeek R1-32B architecture, incorporating native BitNet features for ternary (b1.58-style) support and offering ready-to-run GGUF quantizations. A key differentiator is its updated tokenizer, which includes specialized tags for Routing, Media, Vision, Sound, Tool calls, and Robotics, enabling advanced domain-specific applications.

Key Capabilities

  • Optimized for CPU Inference: Designed for high-quality CPU performance, supporting TQ_2 quantization on Llama.cpp and Ollama via Quantization-Aware Training (QAT).
  • Advanced Tool Calling: Features a unique tokenizer with tags for tool calls, routing, and robotics, enhancing agentic and instruct models for precise function calling, even on smaller models.
  • Cloud-Ready & Cost-Efficient: Positioned as a cloud-ready model that can reduce infrastructure costs, suitable for expert roles in RAG deployments.
  • Specialized Domain Support: Includes specific tags for multimedia, vision, sound, and robotics, making it highly adaptable for niche applications.

Good For

  • Agentic AI Development: Ideal for building sophisticated coding agents, robotics control, and advanced tool-use applications.
  • Resource-Constrained Environments: Excellent choice for deployments where GPU resources are limited, leveraging efficient CPU inference.
  • Domain-Specific Expert Systems: Well-suited for RAG systems where an expert model with specialized understanding of routing, media, or robotics is required.
  • Custom QAT Needs: Offers services for tailored ternary compression and QAT from custom datasets for specific tasks.

Finetunes
1 models