TielCoder 35B-A3B MoE Coding Model Released with Optimized 4-bit Quantization
A new 35B-A3B Mixture of Experts (MoE) coding model named TielCoder has been released, built on top of the Ornith-1.5 fine-tune. Its 22 GB 4-bit quantized version reportedly matches the performance of larger models like Opus4.6 medium on real-world coding tasks while running efficiently on consumer hardware. This release provides local LLM users with a highly efficient, fast, and capable coding assistant that can run on consumer-grade hardware. By leveraging MoE architecture and advanced quantization, it bridges the gap between massive proprietary models and local, resource-constrained setups. TielCoder utilizes a code-weighted importance matrix (imatrix) for dynamic quantization and features a chat template optimized for token-efficient agentic coding. It is available in both GGUF and Apple MLX formats, making it highly accessible for macOS and CPU/GPU setups.
## BACKGROUND
Mixture of Experts (MoE) models contain a large number of total parameters but only activate a small subset (active parameters, like 3B in a 35B model) per token, reducing computational costs during inference. Importance matrix (imatrix) quantization uses calibration data to minimize quality loss when compressing models to lower bit-widths, while Apple's MLX framework optimizes machine learning execution specifically for Apple Silicon.