FIELD GUIDE 01 / GENOMICS

Evo 2 7B

A dense genomic foundation model that stretches to million-base context by interleaving three Hyena filter families with periodic global attention.

Family
StripedHyena 2
Scale
7B dense parameters
Shape
32 layers · width 4,096
Context
Up to 1,048,576 bases

THE 60-SECOND READ

Follow one DNA base.

  1. 01

    A nucleotide becomes a learned vector in a 4,096-wide residual stream.

  2. 02

    Short explicit (Hyena-SE), medium regularized (Hyena-MR), and long implicit (Hyena-LI) operators mix sequence information without full attention at every layer.

  3. 03

    Five multi-head attention (MHA) layers provide periodic global bridges through the 32-layer stack.

  4. 04

    The final vector becomes logits—unnormalized scores—for the next DNA token.

OUTSIDE IN

Whole model

The released 7B checkpoint is a 32-layer, width-4,096 autoregressive model. It was trained on 2.4 trillion tokens; the more-than-8.8-trillion-nucleotide figure describes the broader OpenGenome2 dataset, not the amount consumed by this checkpoint.

INPUTDNA tokensA C G T …
EMBEDWidth 4,096Learned vectors
32 EXACT LAYERS9 SE + 9 MR + 9 LI + 5 MHAThree filter ranges, periodic global bridges
OUTPUTNext-base logitsScores over DNA tokens
How to read this diagram: move left to right from a base token to the exact 32-layer family mix, then to the next-base score. The pipeline is conceptual; the counts are exact for the released 7B 1M configuration.On narrow screens, pan horizontally to inspect the full pipeline.

MECHANISM CUTAWAYS

Open the blocks

01DNA token pathTurn bases into a next-base prediction.

Plain language. Evo 2 reads DNA the way an autoregressive language model reads text: one discrete token at a time, always predicting what follows.

Technical read. A learned embedding places tokens in the width-4,096 residual stream. After 32 sequence- and channel-mixing layers, an output projection produces logits over the tokenizer vocabulary.

Why it matters. The same objective can expose biological regularities because the probability of a base depends on sequence context across motifs, genes, and longer genomic regions.

ACGTembedding4,096 values32 layersP(next base)A .18 · C .07 · G .11 · T .64
How to read this diagram: the highlighted base enters the residual stream; the example probabilities are illustrative, while the width and depth labels are exact.On narrow screens, pan horizontally to follow the full token path.
02Exact 32-layer ribbonSee every mixer in the 1M checkpoint.

Plain language. The model cycles through short, medium, and long filters, then periodically inserts global attention.

Technical read. Zero-based indices 0–31 contain 9 Hyena-SE, 9 Hyena-MR, 9 Hyena-LI, and 5 MHA layers. Attention is sparse in depth—only five layers use it—but each MHA layer itself uses a global 32-head attention map with interpolated Rotary Position Embedding (RoPE).

Why it matters. Most layers keep sequence mixing efficient, while global MHA layers create direct long-range communication checkpoints.

Hyena-SE ×9Hyena-MR ×9Hyena-LI ×9MHA ×5
How to read this diagram: each numbered cell is one real layer, in execution order from 0 to 31. Color and letter labels redundantly identify the four mixer types.On narrow screens, pan horizontally to inspect all 32 layers.
03Hyena operator anatomyProject, mix, filter, gate, return.

Plain language. A Hyena operator makes three views of the input, uses one to control another through a convolutional filter, then gates the result.

Technical read. The conceptual signal flow projects query-like (q), key-like (k), and value-like (v) streams, applies a width-3 short depthwise causal convolution, multiplies k and v, applies the operator-specific inner convolution, gates with q, and output-projects back to the residual stream.

Why it matters. The skeleton stays fixed while the inner filter cartridge changes the range and computational form of sequence mixing.

CONCEPTUAL SIGNAL FLOW

input u
linear projection
q streamk streamv streamwidth-3 depthwise causal convolution
k × v
inner filter
SE / MR / LI
gate × q
output projection
How to read this diagram: follow the projected streams into multiplicative mixing and the replaceable inner filter. This is a teaching abstraction, not an exact tensor graph.On narrow screens, pan horizontally to inspect the full operator flow.
04Replaceable filter cartridgeThree exact filter families, one operator skeleton.

Plain language. The short filter catches local motifs, the medium filter covers a wider explicit window, and the long filter represents very long dependencies compactly.

Technical read. Hyena-SE uses an explicit finite impulse response (FIR) filter of length 7 with 256 groups. Hyena-MR uses length 128 with 256 groups plus an exponential-decay envelope. Hyena-LI uses an implicit exponential/pole-residue filter with 4,096 groups and state size 16.

Why it matters. Hyena-LI can switch from convolutional processing to a compact recurrent form during autoregressive decoding, while SE and MR supply efficient explicit local structure.

hcs / HYENA-SEExplicit FIR
Length
7
Groups
256
Form
short explicit
hcm / HYENA-MRRegularized FIR
Length
128
Groups
256
Envelope
exponential decay
hcl / HYENA-LIImplicit long filter
Groups
4,096
State
16
Decode
recurrent form
How to read this diagram: each card plugs into the same inner-filter slot above. The parameter values are exact for the released Evo 2 7B 1M configuration.
05Long-range viewDifferent spans, plus five global bridges.

Plain language. Several kinds of memory look across the same DNA: local filters see motifs, wider filters see neighborhoods, long implicit filters carry distant structure, and attention occasionally connects everything directly.

Technical read. The range bands below are conceptual, not literal receptive-field boundaries. The five MHA bridge locations correspond to exact layers 3, 10, 17, 24, and 31.

Why it matters. Range specialization avoids paying global-attention cost in every layer while preserving periodic direct access across the full sequence.

CONCEPTUAL RANGE; EXACT MHA LAYERS

DNA sequence · 1,048,576-base configured maximum
SElocal motif span
MRmedium explicit span
LIlong implicit / recurrent span
MHA L3MHA L10MHA L17MHA L24MHA L31
How to read this diagram: longer horizontal marks denote increasing conceptual range; the five labeled bridges mark exact global-attention layers.

READ THE ORIGINALS

Primary sources

The diagrams on this page are original teaching renderings. They simplify layout to make relationships readable; the linked paper, configuration, and implementation remain authoritative for exact computation and checkpoint behavior.