Newsletter image

Subscribe to the Newsletter

Join 10k+ people to get notified about new posts, news and tips.

Do not worry we don't spam!

By pressing the Subscribe button, you confirm that you have read and are agreeing to our Privacy Policy and Terms of Use

Search

GDPR Compliance

We use cookies to ensure you get the best experience on our website. By continuing to use our site, you accept our use of cookies, Privacy Policy, and Terms of Service.

Arc Institute - Biological AI

Evo-2

The Arc Institute and NVIDIA have released Evo-2, a massive open-source genomic AI model designed to advance biological research and applications. Evo-2 is a 40-billion-parameter model trained on 9.3 trillion nucleotides from over 128,000 species, offering unprecedented capabilities in reading, interpreting, and writing genetic code. Developed by a team from Arc, NVIDIA, and other academic institutions, it leverages NVIDIA's powerful GPUs and new architecture for faster data processing. Evo-2 can predict genetic mutations with high accuracy and design genomes, with applications spanning healthcare, agriculture, and industrial sectors. It's fully open-source, available on Arc's GitHub, allowing wide accessibility for research and development. Despite its potential, Evo-2 faces challenges like ethical considerations and high computational demands.
2025-02-19
Updated 2025-03-13 09:24:04

Arc Institute and NVIDIA Drop Evo-2: Biology’s AI Monster Unleashed

The Arc Institute and NVIDIA just yanked the curtain off Evo-2, a biological AI juggernaut. Released today, February 19, 2025, this ain’t some toy model; it’s the biggest open-source genomic AI ever, trained on 9.3 trillion nucleotides from 128,000+ species. Think DNA-reading, genome-writing insanity—Arc and NVIDIA are flexing hard, and it’s free for all. Hype or the real deal? Let’s tear it apart with the dirt straight from the sources.

The Big Drop: What’s Evo-2 Packing?

Evo-2 is a 40-billion-parameter foundation model, cooked up by Arc Institute—a Palo Alto nonprofit hell-bent on cracking biology’s code—and NVIDIA, the GPU kings. It’s trained on a insane dataset: 9.3 trillion nucleotides from bacteria, archaea, phage, humans, plants, and eukaryotes—over 128,000 genomes scraped from the OpenGenome2 atlas. Arc’s site (arcinstitute.org) calls it “the largest publicly available AI model for biology to date.It’s “one of the largest-ever open-source ML models, period”, beating Evo-1’s single-cell focus with a tree-of-life sweep. Hands-on? It crunches sequences up to 1 million nucleotides at once.

The Brains: Who Built This Beast?

Arc’s dream team—Patrick Hsu (UC Berkeley bioengineer, Arc co-founder), Brian Hie (Stanford chem engineer), and a crew from UC Berkeley, Stanford, and UCSF—teamed with NVIDIA’s muscle. Greg Brockman (OpenAI prez) chipped in during a sabbatical, tackling the compute crunch. NVIDIA’s Anthony Costa (digital biology director) and Hani Goodarzi (UCSF gene guru) rounded it out. X’s @pdhsu (5:02 PM CET) says it’s “a universal framework”, Arc’s collab with NVIDIA’s DGX Cloud (blogs.nvidia.com) threw 2,000+ H100 GPUs at it, outmuscling AlphaFold’s FLOPS by 150x. StripedHyena 2, their custom architecture, chows down data 3x faster than transformers—details at asimov.press.

What It Does: DNA Wizardry Unleashed

Evo-2 reads, interprets, and writes genetic code like a punk poet on a bender. It predicts mutations—90% accuracy on BRCA1 breast cancer gene tweaks, per genengnews.com—and designs bacterial-scale genomes from scratch. It “grasps meaning” from 100k+ species’ DNA. Healthcare? Spot disease variants, Agriculture? Engineer badass crops, climate-proof or nutrient-jacked. Industrial? Biofuels or plastic-munching proteins? analyticsindiamag.com says it’s “unimaginable solutions.” Hands-on: Arc’s Evo Designer UI (Arc GitHub) lets you play god, try “design a phage genome” and watch it spit code.

Open-Source Grit: No Paywall BS

Unlike locked-up AI toys, Evo-2’s fully open—weights, training code, inference code, and 8.8 trillion nucleotides of OpenGenome2 data, all on Arc’s GitHub (github.com/arcinstitute) and NVIDIA’s BioNeMo (blogs.nvidia.com). X’s @arcinstitute (3:24 PM CET) flexes: “fully open source” with NVIDIA NIM microservices—tweak it, fine-tune it, break it. Goodfire’s visualizer (Arc collab) shows how it spots codons and introns—rdworldonline.com calls it “transparency done right.” Pinokio fans could run this—16GB VRAM minimum, though.

The Edge: Bigger, Badder, Smarter

Evo-1 was a blurry single-cell snapshot—300 billion nucleotides. Evo-2’s 9.3 trillion dwarf it, hitting 40B params—rivaling Meta’s or DeepMind’s LLMs, says endpts.com. StripedHyena 2 (8x longer context than Evo-1) groks distant genome links—think codons to introns across a million bases. It’s “rewriting multiomics”—drug dev’s drooling over BRCA1 hits, and bioengineers want that AI-designed genome in a lab (UW’s on it).

The Catch: Big Power, Big Risks

It’s not flawless—9.3T nucleotides is massive, but pathogens were axed from training for safety, per aiwire.net. X’s @pdhsu hints at “ethical frameworks” lagging—designing genomes ain’t child’s play. Compute’s a hog—2,000 H100s for months; your rig better be a beast or it’s DOA. Scaling laws hold—more sequences, sharper designs (0.9 AUROC on chromatin)—but it’s a beta beast. Hands-on: Test “predict a protein mutation”—spot the bugs.

Verdict: Evo-2’s Your Bio Punk Dream

Evo-2 isn’t hype—it’s a 40B-parameter monster that reads and writes life like Midjourney forges art. Arc and NVIDIA’s open-source flex—BioNeMo, GitHub, all free—puts it in your hands, not some corp vault. X’s @patrickc says it’s “fundamental”: healthcare, crops, biofuels, chaos. Hands-on: “Design a bacterial genome”

Sources

Arc Institute - Main hub for Evo-2 details.

NVIDIA Blogs - BioNeMo and tech specs.

Asimov Press - Deep dive on genome design.

GENeng News - BRCA1 accuracy and apps.

Analytics India Mag - Healthcare and agri potential.

R&D World Online - Team and transparency.

Endpoints News - Hsu’s vision and scale.

AI Wire - Safety and ethics scoop.

Prev Article
Fino1-8B
Next Article
Baichuan-M1

Related to this topic: