Researchers have published INCAR-SLM, a system combining a fine-tuned small language model with the VASPGuard deterministic post-processor.

The preprint found model size matters less than expected, with performance saturating at a few billion parameters once fine-tuning and post-processing are applied. For reference model sizes with G=16, the system achieves approximately 3 times faster inference than row-independent traversal on Intel processors.

Cloudflare debuted Clairic, an unstructured data extraction model, on 19 October. The company states Clairic processes unstructured data faster, more cheaply, and more accurately than normal large language models.

It appears that both releases are pursuing performance gains without scaling up raw model parameter count.