Release Notes

Summary

This is a stability release. It addresses security vulnerabilities, includes bug fixes, and improves runtime reliability. The 7B and 40B model variants and APIs are unchanged from 2.1.0.

Model Variants

• Evo 2 7B 1M
• Evo 2 40B 1M

Summary

This release introduces support for the Evo 2 7B 1M model variant, providing a faster and more resource-efficient option while maintaining strong accuracy for most use cases. The 40B 1M model remains available as the default for maximum accuracy. This update also includes documentation improvements, clarifies advanced usage and model selection, and updates the release notes to reflect the new model variant.

Summary

This major release introduces support for loading custom and fine-tuned checkpoints, enabling users to leverage their own model weights or community-provided variants. The internal implementation of NIM has migrated to the NVIDIA BioNeMo Framework, providing unified training and inference code bases and standardized checkpoint formats.

Summary

This is a minor update that includes changes that enhance stability for users operating in multi-GPU environments and improve compatibility with new NIM versions:

• Resolved an issue related to FP8 extra states when running on specific combinations of GPU interconnects and CPUs. The fix ensures correct handling of FP8 extra states, with additional checks to verify correct functionality.

• Updated nimlib and related dependencies to improve compatibility with the latest NIM releases.

Summary

This is a minor update that includes the following changes:

• The input prompt is now automatically converted to uppercase internally, helping to prevent the common mistake of supplying lowercase DNA sequences to the model.

• For non-recoverable errors, such as unrecoverable memory ECC errors, the health endpoint will now reflect the error status. This allows cloud infrastructure to redeploy the instance to healthy hardware.

Summary

This is the first release of Evo 2 NIM. Evo 2 is a biological foundation model that is able to integrate information over long genomic sequences, while retaining sensitivity to single-nucleotide changes.

This release features 40 billion parameters model trained on 11 trillion DNA nucleotides with context-length of up to 1 million DNA nucleotides.