Writer Palmyra-mini Family Release

Writer has released the Palmyra-mini family, consisting of three open models ranging from 1.5B to 1.7B parameters. These lightweight models are designed for high performance and efficient inference across a variety of generative and reasoning tasks.

Model Variants and Specializations

The Palmyra-mini family includes three distinct models tailored for different operational needs:

  • palmyra-mini: A non-thinking base model designed as a general-purpose all-rounder for generative tasks.
  • palmyra-mini-thinking-a: A specialized variant optimized for complex reasoning and logic, trained using a Chain of Thought (CoT) approach.
  • palmyra-mini-thinking-b: A specialized variant focused on mathematical equations and advanced problem-solving, also utilizing a Chain of Thought (CoT) approach.

Performance Benchmarks

Each model in the family demonstrates specific strengths across different benchmarks:

  • palmyra-mini achieved a score of 52.6% on Big Bench Hard (get-answer)(exact_match).
  • palmyra-mini-thinking-a achieved 82.87% on GSM8K (strict match) and holds the highest overall average score across benchmarks relative to the other models in the release.
  • palmyra-mini-thinking-b scored 92.5% on AMC23 and holds the highest average benchmark scores for AIME24, AIME25, GPQA, HMMT25, HLE, MMLU_PRO, and MATH500, and LCB relative to the other models in the release.

Technical Implementation and Architecture

All models in the Palmyra-mini family are based on the Qwen architecture, ensuring compatibility with popular inference frameworks including vLLM, SGLang, TRTLLM, and TGI.

Base Model for Thinking-B

For the palmyra-mini-thinking-b model, the base model used was nvidia/OpenReasoning-Nemotron-1.5B. Writer applied Reinforcement Learning (RL) fine-tuning to this base, which improved single-shot accuracy (pass@1) but resulted in a reduction of sampling diversity, leading to a drop in majority@64 performance compared to the SFT base model.

Deployment and Availability

To facilitate ease of deployment, the models are available in multiple formats:

  • Standard weights: Available on Hugging Face.
  • Quantizations: GGUF and MLX-BF16 quantizations are provided for palmyra-mini, palmyra-mini-thinking-a, and palmyra-mini-thinking-b.

Sources