Genomic language models have had a loud few years, and the claims are hard to place against each other unless you follow the field full time. This is an attempt to put all of them on one pair of axes.
Three things the chart makes visible that a list of papers does not:
The frontier is flat, not a staircase. GenSLM reached 25B in October 2022, Evo 2 hit 40B in February 2025, and EDEN came in at 28B in January 2026. Four years, no order of magnitude. Almost everything else since 2025 has been at 1B or below.
The interesting split is tokenization, not size. Reading one base at a time became the default, k-mers came back at 8B in Carbon, and learned tokenization arrived as a serious third option in 2026.
Architectures diversified loudly, then quietly converged. State-space and hybrid models got the attention, but most models here are still plain Transformers.
Exploring it
Hover or tap any point for the detail; click it to open the primary source. Every bar in the lower two panels is a filter: click Hybrid and the scatter narrows to the four hybrid models, while the tokenization panel recounts to show how those four read DNA. One filter is active at a time, and clicking it again clears it.
Reading it
Release dates are the first public preprint or model release, not journal publication. Sizes are the largest reported checkpoint, which for every model except EDEN is also the largest one you can download. GPN-MSA and GPN-Star use whole-genome alignments at training and inference, so they are not directly comparable with single-sequence models.
AlphaGenome, JanusDNA, gLM2, Gene42, TrinityDNA and GENERATOR-v2 are deliberately absent: either the parameter count is not public, or it is not comparable on this axis, or I could not ground it in a primary source. gLM2 is out for a different reason: it is mixed-modality, tokenizing coding regions per amino acid and intergenic regions per nucleotide, so it has no honest home on the colour axis.