Writer open-sources the Palmyra-mini family

Background

Writer's Palmyra line had been built around larger models. With Palmyra-mini, the company released a set of small open models that it describes as complementing the larger Palmyra models for speed-sensitive work.^[1][2][3]

What was published

The family consists of three models, all released in September 2025 as open models. The base model appeared on Hugging Face on 5 September 2025; the two thinking variants, Palmyra-mini-thinking-a and Palmyra-mini-thinking-b, followed on 10 September 2025.^[1][2][3]

All three are released under the Apache 2.0 licence and are labelled English-only.^[1][2][3] Apache 2.0 is a permissive open-source licence: it allows use, modification and redistribution, including inside commercial products, as long as the licence and attribution notices are kept. In practice that means the weights can be picked up without a separate commercial agreement.

Each of the three lists a context window of 131,072 tokens.^[1][2][3] The context window is the amount of text the model can attend to at once — prompt plus generated output. A window of this length corresponds to a long document or a long running conversation rather than a single short question.

Two routes to a reasoning model

The two thinking variants are not two sizes of the same thing. The base model and thinking-a are fine-tuned from Qwen2.5-1.5B, while thinking-b starts from Nvidia's OpenReasoning-Nemotron-1.5B.^[1][2][3] Fine-tuning here means continuing to train an existing checkpoint on additional data instead of training from scratch, so the choice of starting point shapes what the finished model behaves like.

Both thinking variants are presented around mathematical and programming performance, and they differ mainly in their base checkpoints.^[2][3] Offering the same kind of post-training on two different foundations gives users a direct way to compare how much of a result comes from the starting checkpoint rather than from the later training.

Size and distribution formats

The family sits between 1.5 and 1.7 billion parameters, small enough for on-device use.^[1][2][3] Models at that scale can run on a laptop or a phone-class device instead of a data-centre accelerator, which is what makes local and low-latency deployment realistic.

Beyond the original weights, GGUF and MLX builds are published in the same Hugging Face organisation.^[4] GGUF is a file format widely used by local inference tools, and MLX is Apple's framework for machine learning on Apple silicon; publishing in both lets the models load in common local setups without the user doing the conversion work.

Where it fits

The release puts a small open model next to a company's larger models. Writer describes Palmyra-mini as complementing its larger Palmyra models for speed-sensitive work, rather than replacing them.^[1][2][3] For a user, the practical value of a 1.5-billion-parameter model is latency and hardware cost: it answers quickly and runs on hardware people already own.

The limits are stated plainly on the model cards: the models are English-only.^[1][2][3] Teams working in other languages, or on tasks that need a great deal of stored knowledge and long reasoning, would still be looking at the larger models this family is meant to sit alongside. The checkpoints also come from other organisations — Qwen and Nvidia — which is typical for small models, where existing weights are frequently reused as a starting point rather than trained from zero.