CMSManhattan/JiRackUltra_32b

TEXT GENERATIONConcurrent Unit Cost:2Model Size:32.8BQuant:FP8Context Size:32kTool Calling:SupportedPublished:Aug 3, 2026License:otherArchitecture:Transformer Featherless Exclusive Cold

JiRack Ultra 32B is a 32 billion parameter model developed by CMSManhattan, built on a DeepSeek R1-32B architecture with native ternary (BitNet-style) support. Optimized for CPU inference, it features an updated tokenizer with specialized tags for Routing, Media, Vision, Sound, Tool call, and Robotics. This model is designed for efficient cloud deployment and can be used as an expert model in RAG systems, offering various GGUF quantizations for flexible resource utilization.

Loading preview...

JiRack Ultra 32B: CPU-Optimized Ternary Model

JiRack Ultra 32B, developed by CMSManhattan, is a 32 billion parameter model engineered for efficient CPU inference. It is built upon a DeepSeek R1-32B architecture and incorporates native ternary (BitNet-style) features, including a BitLinear ternary path with λ-warmup STE. This design allows for significant compression, with available GGUF quantizations ranging from Q4_K_M (19.5 GB) to Q2_K (13.1 GB), making it suitable for environments with limited memory.

Key Capabilities & Features

  • Ternary Architecture: Native BitNet-style support for enhanced compression and efficiency.
  • Specialized Tokenizer: Features an updated tokenizer with unique tags for Routing, Media, Vision, Sound, Tool call, and Robotics, indicating potential for multimodal and tool-use applications.
  • CPU Optimization: Designed for fast and efficient inference on CPU hardware, reducing cloud infrastructure costs.
  • Cloud-Ready: Positioned as a cloud-ready model, ideal for integration into RAG deployments as an expert model.
  • Flexible Quantizations: Offers various GGUF quantizations (Q4_K_M, Q3_K_M, Q2_K) to balance quality and memory footprint.

Ideal Use Cases

  • Cost-Sensitive Deployments: Excellent for users looking to minimize cloud infrastructure costs due to its CPU optimization.
  • RAG Systems: Can serve as an expert model within Retrieval Augmented Generation (RAG) setups.
  • Resource-Constrained Environments: Suitable for deployment on systems with 24-48 GB RAM, even for the Q4_K_M variant.
  • Specialized Applications: The updated tokenizer with tags for Media, Vision, Sound, Tool call, and Robotics suggests utility in applications requiring these specific functionalities.

Finetunes
1 models