← All models

Marin 8B Instruct

Marin · 2025-05-19 · 8B parameters

Marin 8B Instruct is the supervised fine-tune of Marin 8B Base that the Marin open lab, then run by Stanford CRFM, released on May 19, 2025 under Apache 2.0. It starts from the 13.7T-token Deeper Starling base checkpoint, uses the Llama 3.1 8B architecture with a variant of the Llama 3 tokenizer, and was fine-tuned for about 5.3B tokens on a mix including the Tülu 3 SFT mixture, SmolTalk, OpenThoughts math and Dolphin-R1 data, with no preference tuning or RL. Marin trained it with 4,096-token sequences, the limit its configuration declares.

Benchmark scores