Palmyra-mini

What it is

Palmyra-mini is a small English-language model published by Writer on Hugging Face on 5 September 2025^[1]. Its card lists 1.7 billion parameters and a context window of 131,072 tokens^[1]. A context window is the span of text a model can hold in view at one time; at 131,072 tokens it can take in a long document or an extended conversation without splitting it into pieces. The weights are released under the Apache 2.0 licence, which permits reuse, including commercial reuse, with attribution^[1].

Where it comes from

Palmyra-mini is a fine-tune of Qwen2.5-1.5B^[1]. Fine-tuning starts from a model that has already been trained on a large general corpus and continues training on a narrower, targeted body of data, so that the model's behaviour shifts toward a particular kind of task. Here the base model supplies general English ability, and the fine-tune is aimed at the step-by-step working that mathematics problems require.

What the benchmarks show

The card reports standard benchmark scores. On GSM8K, a set of grade-school mathematics word problems, the model scores 0.818 under strict matching, meaning answers must match the reference exactly rather than being judged approximately correct^[1]. On MATH500, a smaller set of competition-style mathematics problems, it also scores 0.818^[1]. On AMC23, a set of competition problems, it scores 0.6^[1], and on BBH, a collection of deliberately hard reasoning tasks, 0.5259^[1]. On MMLU-Pro, a broad multidisciplinary knowledge test in multiple-choice form, it scores 0.314^[1]; on GPQA diamond, a set of graduate-level science questions written to be difficult for search engines to answer, 0.442^[1]. On the AIME24 problem set it scores 0.2 and on AIME25 0.25^[1]. For code, it scores 0.5 on HumanEval and 0.1519 on LiveCodeBench code generation^[1]. The card gives these results as an unordered list rather than arranging models into a ranking^[1].

Where it sits

Writer frames the release around grade-school mathematics, pointing to GSM8K as the strongest single result^[1]. The pattern in the numbers fits that framing: the score is highest on school-level word problems and falls away on the harder competition sets and on code. A model of this size is small enough to run and adapt cheaply, and the Apache 2.0 licence removes most restrictions on doing so^[1].

These are the scores reported on the model card, and they are not a substitute for trying the model on your own tasks. A benchmark is a fixed set of problems with a fixed scoring rule; a single number says little about how a model handles a real request, how it behaves when it is unsure, or what it does with a long document. Read the table as a rough map of where the model is strong, not as a verdict.