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

JiRack Ultra 32B by CMSManhattan is a 32.8 billion parameter model built on a DeepSeek R1-32B architecture, optimized for efficient CPU inference. It features native ternary (BitNet-style) support and an updated tokenizer with specialized tags for Routing, Media, Vision, Sound, Tool call, and Robotics. This model is designed for cloud-ready deployments, serving as an expert model in RAG systems and excelling in advanced tool calling and robotics applications.

Loading preview...

JiRack Ultra 32B: CPU-Optimized Ternary Model

JiRack Ultra 32B is a 32.8 billion parameter model developed by CMSManhattan, specifically engineered for fast and efficient inference on CPUs. It leverages a DeepSeek R1-32B architecture, incorporating native ternary (BitNet-style) features for enhanced compression and performance. A key differentiator is its updated tokenizer, which includes specialized tags for Routing, Media, Vision, Sound, Tool call, and Robotics, making it highly versatile for domain-specific applications.

Key Capabilities

  • CPU-Optimized Performance: Designed for high-quality CPU inference, supporting GGUF quantizations (Q2_K, Q3_K_M, Q4_K_M) for various memory footprints.
  • Ternary Architecture: Features native BitLinear ternary path (b1.58-style) for efficient model compression and faster execution.
  • Advanced Tool Calling: Utilizes a specialized tokenizer with dedicated tags to enable high-quality tool calling, suitable for agentic and instruct models.
  • Robotics and Routing: Includes specific tokenizer tags for robotics and routing applications, expanding its utility in specialized domains.
  • Cloud-Ready Deployment: Optimized for cost-effective cloud infrastructure, ideal for use as an expert model in RAG deployments.
  • ONNX Export Support: Provides guides for exporting to ONNX for further deployment flexibility.

Good for

  • CPU-constrained environments: When GPU resources are limited or cost-prohibitive.
  • Agentic workflows: Leveraging its advanced tool-calling capabilities for complex automation.
  • Robotics applications: Utilizing its specialized tokenizer for robotic control and understanding.
  • RAG deployments: Serving as an expert model to enhance retrieval-augmented generation systems.
  • Edge computing: Deploying powerful language capabilities on devices with limited hardware.

Finetunes
1 models