CMSManhattan/JiRackUltra_1b

TEXT GENERATIONConcurrent Unit Cost:1Model Size:1.5BQuant:BF16Context Size:32kTool Calling:SupportedPublished:Aug 2, 2026License:otherArchitecture:Transformer Featherless Exclusive Cold

CMSManhattan's JiRack Ultra 1B is a 1.5 billion parameter model built on a redesigned DeepSeek R1 architecture, featuring native ternary (BitNet-style) support for efficient CPU inference. It incorporates an updated tokenizer with specialized tags for Routing, Tool call, and Robotics, making it highly optimized for specific expert tasks and RAG deployments. The model is designed for cost-effective cloud infrastructure use, offering various GGUF quantizations for flexible memory and performance trade-offs.

Loading preview...

JiRack Ultra 1B: CPU-Optimized Ternary Model

JiRack Ultra 1B, developed by CMSManhattan, is a 1.5 billion parameter model engineered for fast and efficient CPU inference. It leverages a redesigned DeepSeek R1 architecture with native ternary (BitNet-style) features, enabling significant compression and reduced memory footprint. The model's tokenizer has been updated to include specialized tags for Routing, Tool call, and Robotics, enhancing its capability for targeted applications.

Key Capabilities & Features

  • CPU-Optimized: Designed for efficient performance on standard CPU hardware, making it suitable for edge and cloud deployments.
  • Ternary Architecture: Incorporates BitNet-style ternary support for high compression, with available GGUF quantizations (Q2_K, Q3_K_M, Q4_K_M) offering various size-performance balances.
  • Specialized Tokenizer: Features unique tags for Routing, Tool call, and Robotics, indicating potential for specialized task execution and integration.
  • Cloud-Ready: Positioned as a cost-saving solution for cloud infrastructure, particularly useful as an expert model in RAG deployments.
  • Flexible Deployment: Provided with Docker images for easy setup and a web UI for interaction.

When to Use This Model

  • Resource-Constrained Environments: Ideal for applications requiring low memory and CPU usage, such as edge devices or cost-sensitive cloud deployments.
  • Specialized Task Execution: Its unique tokenizer tags suggest suitability for tasks involving routing logic, tool invocation, or robotic control.
  • RAG Deployments: Can serve as an efficient expert model within Retrieval Augmented Generation (RAG) systems.
  • Cost-Effective Inference: Designed to minimize operational costs through efficient CPU utilization and small model size.

Finetunes
1 models