Introducing the Palmyra-mini family: Powerful, lightweight, and ready to reason!

Source: HuggingFace Date: 2025-09-11 URL: https://huggingface.co/blog/Writer/announcing-palmyra-mini

Summary

Model release: Writer’s Palmyra-mini family — three 1.5–1.7B Qwen-architecture models: palmyra-mini (general), palmyra-mini-thinking-a (reasoning/logic, GSM8K 82.87%), and palmyra-mini-thinking-b (math, AMC23 92.5%). GGUF and MLX-BF16 quantizations available. Compatible with vLLM, SGLang, TRTLLM, TGI. Notable finding: RL fine-tuning improved pass@1 but reduced sampling diversity (majority@64), indicating mode collapse risk in small reasoning models.

Implications

Thread: open-weights ecosystem health / model release cadence. The AMC23 92.5% result from a 1.7B model is strong for this size class. The RL-fine-tuning tradeoff note is the most honest and useful disclosure: pass@1 up, majority@64 down is exactly the mode collapse signature that makes RL-trained small models less useful for sampling-based applications (agentic use cases that rely on diverse rollouts). Writer releasing GGUF and MLX-BF16 at launch is the right move for the sub-3B model segment, which is dominated by on-device deployments.