Parameter Count
The number of Parameters in a large-language-model (LLM) directly determines model capacity, training complexity, and resource requirements. For example, NVIDIA’s Llama 3.1 Nemotron 70B (70.6 billion parameters) requires ~150GB storage at full precision (32-bit), distributed across 30+ files (~5GB each).
- adam-lucek - quantisation of LLM: Video explaining model-efficiency techniques (e.g., reducing precision from 32-bit to 8-bit), which cuts storage needs by ~75% (e.g., 70B model from ~150GB to ~37.5GB) while preserving model performance.
- model-efficiency: Technique for reducing parameter precision without significant accuracy loss, critical for deploying large models on constrained hardware.
- large-language-model: Model class where parameter count correlates with capability, though architectural efficiency (e.g., MoE) can decouple raw parameter count from performance.
- Ornith-1.0 Agentic Coding LLM Evaluation: 35B MoE Superiority: Evaluation of Ornith-1.0 demonstrates that a 35B Mixture-of-Experts (MoE) model can outperform larger dense models in agentic coding tasks, highlighting that effective parameter utilization and architecture design are as critical as raw parameter count.