Tier 3 Practical open-weight
What the publisher claims: open source, open-source models
The one-liner: An enterprise AI company shipping genuinely permissive small models beside a proprietary platform.
The six questions
Open checkpoints ungated on Hugging Face, including quantized builds.
Apache 2.0 on the open line.
Runs locally, including a full llama.cpp implementation on iPhone.
Training described as business/marketing data focus with synthetic-data techniques; corpus and recipe not published.
Not reproducible from published material.
Public repos and eval methodology are shared, but development direction is entirely Writer's; no community governance.
License
Apache 2.0 (palmyra-base and the palmyra-mini family; verify per release) OSI-approved — license source ↗
- Verify each checkpoint; the enterprise Palmyra X-series and domain models (Med, Fin) are API/proprietary
What is actually released
- weights (ungated on HF)
- GGUF and MLX builds of the mini family
- tokenizer
- evaluation results (lm_eval/lighteval/nemoskills)
- llama.cpp iOS reference implementation
Why it matters
The sponsored-open-source pattern in miniature: a commercial vendor gives away small permissive artifacts that funnel toward a closed product line — and unlike Gemma, the licenses are clean.
Sources
- https://writer.com/blog/palmyra
- https://writer.com/engineering/palmyra-mini-open-source-models
- https://huggingface.co/Writer/palmyra-base
- https://huggingface.co/Writer/palmyra-mini
Derived from the openness dataset behind But Is It Open? — snapshot 2026-11-07 ↗. Scores are qualitative readings of the cited evidence, not verdicts; tier measures openness, not intelligence.