Equimo: Modern Multimodal Models in JAX/Equinox

WARNING: This is a research library implementing recent model architectures. The implementations are based on paper descriptions and may not be exact replicas of the original implementations. Use with caution in production environments.

Equimo provides JAX/Equinox implementations of recent architectures across modalities. Vision is the most complete modality today; language has first-class text encoders/tokenizers, audio includes AST spectrogram models, and tabular includes TabPFN-3 core models.

Features

• Pure JAX/Equinox implementations
• Focus on recent architectures (2023-2026 papers)
• Registry system — register custom attention, FFN, norm, and model classes by name
• BlockChunk — universal building block for staged architectures; supports string-based layer resolution, positional embeddings, downsampling, and stochastic depth
• String-based layer resolution everywhere — pass "layernorm" instead of eqx.nn.LayerNorm
• Modular design for easy experimentation
• Extensive documentation and type hints
• Modality-specific namespaces: equimo.vision, equimo.language, equimo.audio, equimo.tabular
• Generic serialization utilities in equimo.serialization
• Equinox-native fine-tuning utilities in equimo.finetune

Installation

From PyPI

pip install equimo

From Source

git clone https://github.com/clementpoiret/equimo.git
cd equimo
pip install -e .

Package Layout

Equimo is organized by modality, with reusable building blocks separated from modality-specific code:

This is a breaking layout change. The old top-level equimo.models, equimo.layers, equimo.io, equimo.implicit, and equimo.experimental entrypoints are intentionally removed.

Migration Cheat Sheet

Implemented Vision Models

Beyond a standard ViT (e.g., DINOv2 or SigLIP), Equimo provides other SotA architectures:

*: Only contains the Linear Angular Attention module. It is straightforward to build a ViT around it, but may require an additional __call__ kwarg to control the sparse_reg bool.

†: DINOv3 is a VisionTransformer configuration using RoPE positional embeddings and SwiGLU FFN. Pretrained weights are available — see pretrained models.

‡: FreeNet building blocks (FreeNetBlock, S2Mixer, ShiftNeck) are implemented in equimo.vision.layers and registered in the convolution registry. There is no standalone FreeNet model class; use BlockChunk to compose a full network from these blocks.

Implemented Audio Models

Implemented Tabular Models

Equimo implements the core TabPFN-3 classifier and regressor architecture in JAX/Equinox, with Equimo-native converted weights. TabPFN-3 is a transformer foundation model from Prior Labs that uses in-context learning to solve tabular prediction tasks in a single forward pass. The upstream model card describes it as intended for structured classification and regression tasks up to 1M samples and 2000 features, trained purely on synthetic tabular tasks.

Equimo exposes the model core directly: inputs are unbatched x, y, and n_train arrays, not the upstream TabPFNClassifier/TabPFNRegressor sklearn-style preprocessing and ensembling API. Classification variants return test-row log probabilities over the class vocabulary; regression variants return raw 5000-bucket logits.

Available constructors:

• Classifier: tabpfn(), tabpfn_v3_classifier_default(), tabpfn_v3_classifier_binary(), tabpfn_v3_classifier_multiclass(), tabpfn_v3_classifier_ood()
• Regressor: tabpfn_regressor(), tabpfn_v3_regressor_default(), tabpfn_v3_regressor_mediumdata(), tabpfn_v3_regressor_ood(), tabpfn_v3_regressor_timeseries()

The specialized checkpoint names mirror the Prior Labs release: binary and multiclass classifier variants, OOD classifier/regressor variants, a medium-data regressor, and a time-series regressor. Upstream TabPFN-3 weights are released under tabpfn-3-license-v1.0; review that license before using pretrained weights outside research or internal evaluation.

Vision Usage

import jax
import equimo.vision.models as em

key = jax.random.PRNGKey(0)
model = em.VisionTransformer(
    img_size=224,
    in_channels=3,
    dim=384,
    patch_size=14,
    num_heads=[6],
    depths=[12],
    num_classes=1000,
    key=key,
)

x = jax.random.normal(key, (3, 224, 224))

# Inference (dropout disabled)
logits = model(x, key=key, inference=True)

# Feature extraction
features = model.features(x, key=key, inference=True)

Examples

See docs/usage.md for a compact non-fine-tuning usage guide covering model construction, feature extraction, text encoders, TabPFN, serialization, and registries.

Runnable examples live under examples/:

Fine-Tuning

equimo.finetune provides Equinox-native model-side fine-tuning primitives: trainability masks, parameter labels, LLRD metadata, head replacement, feature pooling, LoRA, adapters, prompts, scale/shift, IA3, DoRA, model deltas, and model-merging utilities.

Equimo intentionally does not provide optimizers, schedules, dataloaders, or training loops. The core abstraction is a FineTunePlan, which partitions a model into trainable and frozen PyTrees and provides labels/group metadata for external optimizers such as Optax or Rollfast.

Minimal example:

import jax
import equimo.finetune as eqft
import equimo.vision.models as em

key = jax.random.PRNGKey(0)
model = em.vit_tiny_patch16_224(num_classes=10, key=key)

plan = eqft.prepare_finetune(
    model,
    trainable=eqft.TrainableSpec(
        mode="full",
        freeze=eqft.TargetSpec(tags_any=("embedding.patch",)),
    ),
    labels=eqft.LLRDConfig(decay=0.75),
)

trainable = plan.trainable
frozen = plan.frozen

def loss_fn(trainable, batch):
    model = plan.combine(trainable)
    logits = jax.vmap(lambda x: model(x, key=key, inference=False))(batch["x"])
    return cross_entropy(logits, batch["y"])  # supplied by user code

Frozen leaves are absent from plan.trainable; they are not assigned a zero learning rate. plan.labels and plan.group_specs are ready for an external optimizer partition.

Rollfast example:

import rollfast.finetune as rfft

optim = rfft.adamw_from_plan(
    plan,
    total_steps=20_000,
    base_lr=5e-4,
    schedule="warmup_cosine",
    weight_decay=0.05,
    clip_global_norm=1.0,
    accumulation_steps=4,
)
opt_state = optim.init(plan.trainable)

For Rollfast Schedule-Free optimizers, combine optim.eval_params(trainable, opt_state) with plan.frozen for validation or checkpointing. For EMA/SWA, request named views such as view="ema" or view="swa" from the same method. For optimizer-state memory savings, rfft.adamw8_from_plan can quantize large, eligible AdamW moment groups while Equimo still emits the same plan metadata. For sharpness-aware updates, rfft.make_sam_step, rfft.SAMConfig, and rfft.ASAMConfig wrap a Rollfast base optimizer with an explicit two-pass SAM/ASAM step. For AdaLoRA-style runs, rfft.make_adalora_controller emits fixed-shape rank support masks for Equimo AdaLoRA adapters; use eqft.lora_rank_groups to build controller groups and eqft.apply_lora_rank_pattern to apply emitted masks. For LP-FT or gradual unfreezing, rfft.reconfigure_optimizer can migrate compatible optimizer state between two Equimo plans and report what changed. After optimizer initialization, rfft.optimizer_state_memory_summary reports measured state bytes, including 8-bit state and Kron preconditioner factors. Before initialization, rfft.estimate_optimizer_state_memory can estimate optimizer-family moment state and Kron preconditioner factors from the Equimo plan without materializing optimizer state. Structured Rollfast optimizers use the same plan: rfft.hybrid_aurora_adam_from_plan, rfft.hybrid_prism_adam_from_plan, and rfft.hybrid_kron_adam_from_plan compile Aurora/PRISM/Kron groups without adding a Rollfast dependency to Equimo core. Use rfft.make_state_checkpoint(...) for Rollfast optimizer state; Equimo model and delta serialization remain separate.

PEFT example:

lora_model = eqft.apply_lora(
    model,
    eqft.LoRAConfig(rank=8, alpha=16.0),
    key=key,
)
lora_plan = eqft.prepare_finetune(
    lora_model,
    trainable=eqft.TrainableSpec(mode="peft", method_name="lora"),
)

eqft.save_delta(lora_model, "my_lora.eqft", method="lora")

See docs/finetuning for selectors, linear probing, LP-FT, LLRD, PEFT methods, serialization, and Optax/Rollfast integration examples.

Predefined Model Variants

Every model family ships convenience constructors that encode the canonical hyperparameters for each published variant. Under the hood each constructor resolves through a two-level registry: a shared base config (e.g. in_channels=3) and a variant config (depths, widths, …) that are merged before the model is instantiated. Any key can be overridden at call time via **kwargs.

import jax
import equimo.vision.models as em

key = jax.random.PRNGKey(0)

# Build a predefined variant with default config
model = em.vit_base_patch16_224(key=key)

# Override specific parameters — e.g. fine-tune head size or disable the class token
model = em.vit_small_patch16_224(num_classes=10, key=key)

# Models with pretrained weights accept a `pretrained` flag
model = em.dinov2_vitb14(pretrained=True)

Available variants

LowFormer requires attention_type ("softmax" or "sigmoid") which has no sensible default and must be supplied by the caller.

Extending the variant registry

Each family exposes its internal registry dict and _build_* function. You can add your own variants without subclassing:

from equimo.vision.models.attnet import _ATTNET_REGISTRY, _ATTNET_BASE_CFG, _build_attnet

_ATTNET_REGISTRY["attnet_custom"] = (
    _ATTNET_BASE_CFG,
    {"depths": [3, 3, 9, 3], "dims": [56, 112, 224, 448], "drop_path_rate": 0.15},
)

model = _build_attnet("attnet_custom", key=jax.random.PRNGKey(0))

Registry System

Equimo exposes registries for layer families and full model classes. Each layer registry follows the same pattern: a register_* decorator and a get_* resolver. All model and layer constructors accept string names wherever a class would normally be passed.

Available registries

equimo.vision.layers.get_layer resolves a string name across core and vision layer registries in priority order, so BlockChunk and vision model constructors can accept a single string for any layer type. equimo.core.layers.get_layer is available for shared/core-only code.

Model registration is modality-aware:

from equimo.registry import get_model_cls, register_model

@register_model("mynet", modality="vision")
class MyVisionModel(eqx.Module):
    ...

assert get_model_cls("mynet", modality="vision") is MyVisionModel

If a name exists in more than one modality, pass modality= to disambiguate.

Example: register a custom attention and build a ViT with it

import jax
import jax.numpy as jnp
import jax.random as jr
import equinox as eqx
from jaxtyping import Array, Float, PRNGKeyArray

import equimo.vision.models as em
from equimo.vision.layers import register_attn, register_attn_block

# ── 1. Define and register the attention module ───────────────────────────────

@register_attn("myattn")
class MyAttention(eqx.Module):
    """Minimal scaled dot-product attention (single-head demo)."""

    qkv: eqx.nn.Linear
    proj: eqx.nn.Linear
    dim: int = eqx.field(static=True)

    def __init__(self, dim: int, *, key: PRNGKeyArray, **kwargs):
        self.dim = dim
        k1, k2 = jr.split(key)
        self.qkv = eqx.nn.Linear(dim, 3 * dim, use_bias=False, key=k1)
        self.proj = eqx.nn.Linear(dim, dim, use_bias=False, key=k2)

    def __call__(
        self,
        x: Float[Array, "seq dim"],
        *,
        key: PRNGKeyArray,
        inference: bool = False,
    ) -> Float[Array, "seq dim"]:
        seq, d = x.shape
        qkv = jax.vmap(self.qkv)(x)              # (seq, 3*dim)
        q, k, v = jnp.split(qkv, 3, axis=-1)     # each (seq, dim)
        scale = d ** -0.5
        attn = jax.nn.softmax(
            (q @ k.T * scale).astype(jnp.float32), axis=-1
        ).astype(x.dtype)
        return jax.vmap(self.proj)(attn @ v)


# ── 2. Wrap it in a transformer block and register it ────────────────────────

@register_attn_block("myattnblock")
class MyAttentionBlock(eqx.Module):
    norm1: eqx.nn.LayerNorm
    norm2: eqx.nn.LayerNorm
    attn: MyAttention
    mlp: eqx.nn.MLP

    def __init__(
        self,
        dim: int,
        num_heads: int,   # accepted for API compatibility; ignored here
        mlp_ratio: float = 4.0,
        drop_path: float = 0.0,
        *,
        key: PRNGKeyArray,
        **kwargs,
    ):
        k1, k2 = jr.split(key)
        self.norm1 = eqx.nn.LayerNorm(dim)
        self.norm2 = eqx.nn.LayerNorm(dim)
        self.attn = MyAttention(dim=dim, key=k1)
        self.mlp = eqx.nn.MLP(
            in_size=dim,
            out_size=dim,
            width_size=int(dim * mlp_ratio),
            depth=1,
            key=k2,
        )

    def __call__(
        self,
        x: Float[Array, "seq dim"],
        *,
        key: PRNGKeyArray,
        inference: bool = False,
        **kwargs,
    ) -> Float[Array, "seq dim"]:
        x = x + self.attn(
            jax.vmap(self.norm1)(x), key=key, inference=inference
        )
        x = x + jax.vmap(self.mlp)(jax.vmap(self.norm2)(x))
        return x


# ── 3. Plug it into VisionTransformer via its string name ────────────────────

key = jr.PRNGKey(0)
model = em.VisionTransformer(
    img_size=224,
    in_channels=3,
    dim=384,
    patch_size=16,
    num_heads=[6],
    depths=[6],
    num_classes=1000,
    block="myattnblock",   # ← resolved from the registry
    key=key,
)

x = jax.random.normal(key, (3, 224, 224))
logits = model(x, key=key, inference=True)
print(logits.shape)  # (1000,)

Re-registering an existing name raises a ValueError by default. Pass force=True to override:

@register_attn("myattn", force=True)
class MyImprovedAttention(eqx.Module):
    ...

BlockChunk

BlockChunk (from equimo.vision.layers or equimo.core.layers) is the canonical building block for multi-stage vision architectures. It groups a sequence of identical blocks with optional positional embedding and downsampling, and handles stochastic depth scheduling automatically.

from equimo.vision.layers import BlockChunk
from equimo.vision.layers.attention import AttentionBlock
from equimo.vision.layers.downsample import ConvNormDownsampler
import jax.random as jr

key = jr.PRNGKey(0)

stage = BlockChunk(
    depth=4,
    in_channels=96,
    out_channels=192,
    module="attentionblock",       # resolved from _ATTN_BLOCK_REGISTRY
    module_kwargs={"dim": 96, "num_heads": 3, "mlp_ratio": 4.0},
    downsampler="convnormdownsampler",  # resolved from _DOWNSAMPLER_REGISTRY
    downsampler_kwargs={},         # in_channels/out_channels injected automatically
    downsample_last=True,          # blocks run first, then downsample
    drop_path=0.1,
    key=key,
)

Passing a list of drop-path rates of length depth applies them per block. Any list-valued entry in module_kwargs whose length equals depth is also spread across blocks (e.g. per-block attention types).

Language Usage

equimo.language provides text encoders and tokenizers. Text tokenization relies on tensorflow_text; install Equimo with the language extra:

pip install equimo[language]

Zero-shot classification example using TIPS:

import jax
from einops import rearrange

from equimo.language import SentencePieceTokenizer, TextTransformerEncoder
from equimo.serialization import load_weights
from equimo.vision.io import load_image
from equimo.vision.models import tips_vits14_hr
from equimo.utils import PCAVisualizer, normalize, plot_image_and_feature_map

key = jax.random.PRNGKey(42)
image = load_image("./demo.jpg", size=448)
text = [
    "A baby discovering happiness",
    "A computer",
]

image_encoder = tips_vits14_hr(pretrained=True)
text_encoder = TextTransformerEncoder(
    dim=384,
    mlp_ratio=4.0,
    depth=12,
    num_heads=6,
    vocab_size=32000,
    scale_sqrt_depth=True,
    act_layer="relu",
    temperature=0.005497702397406101,
    key=key,
)
text_encoder = load_weights(text_encoder, identifier="tips_vits14_hr_text")

ids, paddings = SentencePieceTokenizer(identifier="sentencepiece_tips").encode(
    text, max_length=64
)

text_embedding = normalize(
    jax.vmap(text_encoder, in_axes=(0, 0, None))(ids, paddings, key)
)
image_embedding = jax.vmap(image_encoder.norm)(image_encoder.features(image, key))
cls_token = normalize(image_embedding[0])
spatial_features = rearrange(
    image_embedding[2:], "(h w) d -> h w d", h=int(448 / 14), w=int(448 / 14)
)

cos_sim = jax.nn.softmax(
    ((cls_token[None, :] @ text_embedding.T) / text_encoder.temperature), axis=-1
)

label_idxs = jax.numpy.argmax(cos_sim, axis=-1)
cos_sim_max = jax.numpy.max(cos_sim, axis=-1)
label_predicted = text[label_idxs[0]]
similarity = cos_sim_max[0]
pca_obj = PCAVisualizer(spatial_features)
image_pca = pca_obj(spatial_features)

plot_image_and_feature_map(
    image.transpose(1, 2, 0),
    image_pca,
    "./out.png",
    "Input Image",
    f"{label_predicted}, prob: {similarity * 100:.2f}%",
)

Resulting in such a wonderful result:

Saving and Loading Models

Equimo provides utilities for saving models locally and loading pre-trained models from the official repository.

Saving Models Locally

from pathlib import Path
from equimo.serialization import save_model

# Save model with compression (creates .tar.lz4 file)
save_model(
    Path("path/to/save/model"),
    model,
    model_config,
    torch_hub_cfg,  # can be an empty list; used to track weight provenance
    compression=True,
)

# Save model without compression (creates directory)
save_model(
    Path("path/to/save/model"),
    model,
    model_config,
    torch_hub_cfg,
    compression=False,
)

Loading Weights

from equimo.serialization import load_weights
from equimo.vision.models import dinov2_vits14_reg, siglip2_vitb16_256

# Load a pre-trained vision model from the official repository.
model = dinov2_vits14_reg(pretrained=True, dynamic_img_size=True)

# Load a local model (compressed).
model = dinov2_vits14_reg(pretrained=False, dynamic_img_size=True)
model = load_weights(model, path=Path("path/to/model.tar.lz4"))

# Load a local model (uncompressed directory).
model = dinov2_vits14_reg(pretrained=False, dynamic_img_size=True)
model = load_weights(model, path=Path("path/to/model/"))

Constructor parameters are controlled when building the target model:

model = siglip2_vitb16_256(
    pretrained=False,
    dynamic_img_size=True,  # forwarded to VisionTransformer.__init__
)
model = load_weights(model, identifier="siglip2_vitb16_256")

Custom models are restored the same way:

class MyNet(eqx.Module):
    ...

model = MyNet(..., key=key)
model = load_weights(model, path=Path("mynet.tar.lz4"))

List of Pretrained Models

The following models have pretrained weights available in Equimo:

• DINOv2
• DINOv3
• SigLIP2
• TIPS
• EUPE (both ViT and ConvNeXt variants)
• AST
• TabPFN-3

Model identifiers map to filenames in Equimo's HuggingFace repository.

Examples:

• dinov2_vitb14
• dinov2_vits14_reg
• dinov3_vits16_pretrain_lvd1689m
• dinov3_vitb16_pretrain_lvd1689m
• dinov3_vitl16_pretrain_lvd1689m
• dinov3_vits16plus_pretrain_lvd1689m
• dinov3_vith16plus_pretrain_lvd1689m
• dinov3_vit7b16_pretrain_lvd1689m
• dinov3_vitl16_pretrain_sat493m
• dinov3_vit7b16_pretrain_sat493m
• siglip2_vitl16_512
• siglip2_vitso400m16_384
• tips_vitg14_lr
• ast_base_patch16_audioset_10_10_0_4593
• ast_base_patch16_speechcommands_v2_10_10_0_9812
• tabpfn_v3_classifier_default
• tabpfn_v3_classifier_binary
• tabpfn_v3_classifier_multiclass
• tabpfn_v3_classifier_ood
• tabpfn_v3_regressor_default
• tabpfn_v3_regressor_mediumdata
• tabpfn_v3_regressor_ood
• tabpfn_v3_regressor_timeseries

Audio

equimo.audio provides AST spectrogram models and reusable audio layers. AST inputs are single log-mel spectrograms shaped (time, frequency), and batch inference can be done with jax.vmap.

import jax
import jax.numpy as jnp
import equimo.audio.models as am

key = jax.random.PRNGKey(0)
model = am.ast_base_patch16_audioset_10_10_0_4593(pretrained=True)

x = jnp.ones((1024, 128))  # (time, frequency)
logits = model(x, key=key, inference=True)
features = model.forward_features(x, key=key, inference=True)

Pretrained AST checkpoints currently available:

• ast_base_patch16_audioset_10_10_0_4593: Full AudioSet, 10x10 strides, weight-averaged checkpoint.
• ast_base_patch16_speechcommands_v2_10_10_0_9812: SpeechCommands V2-35, 10x10 strides, non-averaged checkpoint.

Tabular

equimo.tabular provides TabPFN-3 models and reusable tabular layers.

import jax
import jax.numpy as jnp
import equimo.tabular.models as tm

key = jax.random.PRNGKey(42)
model = tm.tabpfn_v3_classifier_default(pretrained=True)

x = jnp.ones((12, 5))  # (rows, columns)
y = jnp.array([0, 1, 0, 1, 0, 1, 0, 1, 0, 0, 0, 0])
n_train = 8

log_probs = model(x, y, n_train, key=key, inference=True)
features = model.forward_features(x, y, n_train, key=key, inference=True)

For regression, use tm.tabpfn_v3_regressor_default(pretrained=True) or one of the specialized regressor variants. The regressor returns raw bucket logits; any post-processing to scalar predictions should follow the chosen downstream regression decoding strategy.

Mixed Precision

Equimo follows a strict WYSIWYG policy — modules never silently cast inputs or weights. Cast your model before running inference:

import jax
import jax.numpy as jnp
import equinox as eqx

model_bf16 = jax.tree_util.tree_map(
    lambda leaf: leaf.astype(jnp.bfloat16) if eqx.is_inexact_array(leaf) else leaf,
    model_fp32,
)

Isolated float32 upcasts are mandatory for numerically sensitive operations (softmax, layer norm variance). These are applied internally where needed.

Contributing

Contributions are welcome! Please feel free to submit a Pull Request. For major changes, please open an issue first to discuss what you would like to change.

License

This project is licensed under the MIT License - see the LICENSE file for details.

Citation

If you use Equimo in your research, please cite:

@software{equimo2024,
  author = {Clément POIRET},
  title = {Equimo: Modern Multimodal Models in JAX/Equinox},
  year = {2024},
  publisher = {GitHub},
  url = {https://github.com/clementpoiret/equimo}
}

If you use TabPFN models through Equimo, please also cite the relevant TabPFN work:

@misc{grinsztajn2026tabpfn3technicalreport,
  title = {TabPFN-3: Technical Report},
  author = {L{\'e}o Grinsztajn and Klemens Fl{\"o}ge and Oscar Key and Felix Birkel and Philipp Jund and Brendan Roof and Mihir Manium and Shi Bin and Hoo and Magnus B{\"u}hler and Anurag Garg and Dominik Safaric and Jake Robertson and Benjamin J{\"a}ger and Simone Alessi and Adrian Hayler and Vladyslav Moroshan and Lennart Purucker and Philipp Singer and Alan Arazi and Julien Siems and Jan Hendrik Metzen and Georg Grab and Nick Erickson and Siyuan Guo and Eliott Kalfon and Simon Bing and David Salinas and Clara Cornu and Lilly Charlotte Wehrhahn and Diana Kriuchkova and Kursat Kaya and Lydia Sidhoum and Marie Salmon and Jerry Chen and Madelon Hulsebos and Yann LeCun and Samuel M{\"u}ller and Bernhard Sch{\"o}lkopf and Sauraj Gambhir and Noah Hollmann and Frank Hutter},
  year = {2026},
  eprint = {2605.13986},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2605.13986},
}

@inproceedings{hollmann2023tabpfn,
  title = {TabPFN: A Transformer That Solves Small Tabular Classification Problems in a Second},
  author = {Noah Hollmann and Samuel M{\"u}ller and Katharina Eggensperger and Frank Hutter},
  booktitle = {International Conference on Learning Representations},
  year = {2023},
  url = {https://arxiv.org/abs/2207.01848},
}