Question about Evo2-1B performance across tasks (vs. Evo2-7B / Evo1.5-7B with 8K context)

[JoshPhd]

Hi, thanks for releasing Evo 2 — amazing work!

Because my GPU resources are limited, I’m very interested in whether the 1B model is a practical alternative to the larger Evo2-7B or Evo1.5-7B models when using the same 8K context window.

How well Evo2-1B performs across typical tasks, such as mutation-effect prediction, regulatory sequence modeling, or short-range sequence generation?

[garykbrixi]

Hi @JoshPhd, the 1B model is weaker in most tasks so we consider it a more experimental checkpoint, useful for experiments before scaling up, for certain types of tasks, and for higher throughput.

See others' work in TraitGym which finds the 1B is significantly weaker for variant effect, while others found Evo 2 1B has strong performance on RNA tasks, within a few percent to the best models.

[JoshPhd]

@garykbrixi thx for your reply