Writer drops 1.5B reasoning models that beat benchmarks, not your GPU

Writer just dropped a trio of tiny open models that punch well above their weight class. The Palmyra-mini family consists of three variants—a base model and two specialized ‘thinking’ models—all sitting between 1.5B and 1.7B parameters. That’s small enough to run inference on a laptop, but the benchmark numbers tell a different story entirely.

The base model, palmyra-mini, hits a respectable 52.6% on Big Bench Hard, positioning it as a solid generalist. But the real action is in the two Chain of Thought variants. The palmyra-mini-thinking-a model scores an eyebrow-raising 82.87% on GSM8K, a math reasoning benchmark that trips up models five times its size. Not to be outdone, the thinking-b variant—built on top of Nvidia’s OpenReasoning-Nemotron-1.5B—posts a 92.5% on AMC23 and leads the pack on a slate of brutal tests including AIME24, MATH500, and GPQA.

Writer’s release notes include an unusually candid observation about a trade-off they hit during reinforcement learning fine-tuning. While RL boosted single-shot accuracy (pass@1), it killed sampling diversity, causing a performance drop on majority@64 evaluations. ‘This highlights a trade-off between accuracy and diversity,’ the team notes, flagging mode collapse as an area ripe for further research. That level of transparency is rare in model releases and frankly, refreshing.

All three models use Qwen architecture, which means they’ll drop right into existing inference frameworks like vLLM and SGLang without much fuss. GGUF and MLX quantizations are already available. For developers who’ve been watching the reasoning-model race from the sidelines because they don’t have access to a server farm, this release lowers the barrier considerably. The question now isn’t whether small models can reason—it’s what people will actually build with them.