Open models
Published on Hugging Face under Mohaaxa. Every released quantization ships with a model card covering calibration data, method and measured trade-offs, the same standard the research holds benchmarks to.
- Qwen2.5-1.5B, GPTQ 4-bitQuantized Qwen2.5-1.5B-Instruct. Model card documents calibration data and method.
- Qwen2.5-1.5B, AWQ 4-bitQuantized Qwen2.5-1.5B-Instruct. Model card documents calibration data and method.
- QuantBench leaderboard dataThe full sweep behind R3, failed runs included.
Related: R3, QuantBench and F-03, Vision on 1-bit LLMs.