Palmyra-mini-thinking-b

What it is

Palmyra-mini-thinking-b is a language model published by Writer on Hugging Face on 10 September 2025 ^[1]. It has 1.7 billion parameters ^[1]. The "thinking" in its name marks it as the kind of model that writes out intermediate reasoning — working a problem through in steps — before giving a final answer, and the benchmarks its card reports are the ones such models are usually measured on: competition mathematics, Olympiad-style problems and programming contests ^[1]. The card labels the model as English-only ^[1] and places it under the Apache 2.0 licence ^[1], a permissive open licence that allows reuse and modification.

Where it comes from

The model is not trained from scratch. It is a fine-tune of Nvidia's OpenReasoning-Nemotron-1.5B ^[1], and the released checkpoint is larger than the one it was built from: 1.7 billion parameters against the base model's 1.5 billion ^[1]. Fine-tuning means continuing to train an existing model on a narrower body of material so that it picks up a particular skill or style, rather than teaching it language from zero. It accepts a context window of 131,072 tokens ^[1]. A context window is how much text, counted in tokens, the model can keep in view at once; a token is roughly a word fragment, so a window of this size is on the order of a long document — enough room for a lengthy problem statement together with a long chain of intermediate steps.

The benchmarks it reports

The card is dominated by mathematics and code evaluation. On AMC23, a set of American Mathematics Competitions problems, the model scores 92.50 percent ^[1]. On MATH500 it reaches 88.20 percent at pass@1 and 93.80 percent averaged over 64 samples ^[1]. On AIME24 it scores 59.43 percent and on AIME25 49.69 percent, both averaged over 64 samples ^[1]. Knowledge-oriented tests sit lower: MMLU-Pro at 55.49 percent and GPQA at 42.01 percent, again over 64 samples ^[1]. For code it reports 63.43 percent on Codeforces and 57.33 percent on OlympiadBench ^[1]. HumanEval, a code-generation benchmark, is given as 7.32 percent at pass@1 with a single sample ^[1].

Reading the numbers: one sample and sixty-four

Sampling used temperature 0.6 and top-p 0.95 ^[1]. Temperature controls how much randomness is injected when the model picks each next token, and top-p restricts the choice to the most probable tokens whose combined probability reaches a set share; together the two settings fix how varied a set of answers will be. The card reports both single-sample and 64-sample averages because the two differ widely on hard benchmarks ^[1]. The difference is easy to see on MATH500: a first attempt, pass@1, scores 88.20 percent, while the average over 64 attempts is 93.80 percent ^[1]. On the harder AIME sets the card gives only the 64-sample averages ^[1]. HumanEval's 7.32 percent is a single-sample pass@1 figure ^[1], so it belongs to the stricter column of the card rather than the 64-sample averages.

Where it sits

Small models with open licences have become a practical way to run reasoning work without depending on a hosted service, and this release belongs to that group: a few billion parameters, a long context window, a permissive licence and published benchmark numbers ^[1]. Its card reads as a map of where the model is strong and where it is not. Competition mathematics is clearly the strong side, with AMC23 above 90 percent and MATH500 in the high eighties at a single attempt ^[1]. Broad factual and scientific knowledge is weaker, at 55.49 percent on MMLU-Pro and 42.01 percent on GPQA ^[1]. Code is uneven: a Codeforces score of 63.43 percent ^[1] alongside a single-sample HumanEval score of 7.32 percent ^[1]. Because the card reports both single-sample and 64-sample averages, a reader can see how much of the performance depends on being allowed several attempts — a distinction that matters when the model is used for a task where only one answer will be produced.