Prism ML
Prism ML is an entity focused on developing compact, high-performance AI models optimized for local deployment.
Key Projects
Bonzai 2.7B
A compact language model derived from the qwen-38-27b architecture. It is designed to address performance challenges on single-GPU setups, aiming to improve local AI accessibility.
- Architecture: Derived from Qwen 3.8 27B.
- Goal: Enable powerful local AI access on constrained hardware.
- Performance: Evaluated for single-GPU viability.
- Analysis: See Bonzai 2.7B AI: Single-GPU Performance Challenges for Local AI Accessibility for detailed testing results.
Ternary Bonsai 2 27B
A new 27B-class reasoning model utilizing ternary transformer weights for extreme quantization (1-bit/2-bit quantization).
- Quantization Focus: Recent evaluations specifically target Q1 and Q2 quantized versions to balance memory usage and reasoning performance.
- Benchmarking: Comprehensive re-evaluation of performance, memory footprint, and reasoning capabilities in constrained environments (e.g., 16GB VRAM setups).
- Detailed Analysis: See Q2 Re-evaluation: Benchmarking Performance, Memory, Reasoning for detailed testing results.