Palmyra-mini family released: 1.5B–1.7B models with CoT reasoning and GGUF/MLX quantization

AI Impact Summary

WRITER releases the Palmyra-mini family of 1.5B–1.7B parameter models with Chain-of-Thought reasoning variants. The palmyra-mini-thinking-a and -thinking-b variants show strong benchmark results (GSM8K 82.87%, AMC23 92.5%), indicating meaningful reasoning improvements at a small scale. The release includes GGUF and MLX quantizations and claims compatibility with popular inference stacks (vLLM, SGLang, TRTLLM, TGI); palmyra-thinking-b is built on OpenReasoning-Nemotron-1.5B with RL fine-tuning, noting a trade-off between accuracy and diversity. Organizations should assess CoT-based workloads and framework compatibility when integrating these models into existing pipelines, and plan migration and evaluation accordingly.