crazedcodernate/bioclip-2.5-mobile-fastvit

BioCLIP 2.5 Mobile (FastViT student)

A 23.8 MB on-device plant identification model, distilled from BioCLIP 2.5 Huge (ViT-H/14, roughly 632M parameters).

Author: Nate Hamilton (CrazedCoderNate) Training code: https://github.com/CrazedCoderNate/bioclip-mobile-distill

The student produces embeddings in the teacher's own 1024-dimensional joint image and text space. Species lookup is therefore a single matrix multiply against a precomputed table of text embeddings, which means the text encoder never needs to run on the device. A phone runs the image encoder only.

What it is for

Fast, offline, first-pass plant identification on mobile hardware, in front of a larger model that arbitrates. It is not a replacement for a full-size classifier, and the Limitations section below is not boilerplate.

Results

Evaluated against the teacher on 2,000 held-out images, classifying across all 4,271 species in the lookup table.

metric	value
top-1 agreement with teacher	71.7%
top-5 contains teacher's top-1	88.8%
embedding cosine vs teacher	0.8383
random baseline	0.023%

Model size and precision, measured by embedding cosine against the fp32 export on real images:

build	size	cosine vs fp32
fp32 ONNX	47.0 MB	1.0000 (reference)
fp16 ONNX (recommended)	23.8 MB	1.0000
int8, percentile calibration	12.9 MB	0.7897
int8, entropy calibration	12.9 MB	0.3781
int8, minmax calibration	12.9 MB	0.3781

int8 is not published because it does not work here. FastViT's stage 3 uses self-attention with LayerNorm, and no calibration method tested kept the embedding usable. Ship fp16.

Training

Teacher	BioCLIP 2.5 ViT-H/14, 1024-d
Student	fastvit_sa12 (timm), 11.6M params, ImageNet warm start
Objective	cosine distillation to cached teacher embeddings, plus 0.1 weight smooth L1
Data	iNaturalist 2021 train_mini, kingdom Plantae only
Images	213,550
Species	4,271
Epochs	30, batch 64, AdamW, OneCycleLR
Hardware	one RTX 5070 (12 GB)
Wall clock	30 min teacher embedding pass, 3 hours student training

No labels were used. The teacher's cached embeddings are the entire supervision signal.

Final training cosine was 0.9014 against 0.8412 validation, so the student does not fully match the teacher even on data it trained on. Capacity is binding, and fastvit_sa24 is the obvious next step for anyone wanting more.

Limitations

Do not use this for toxicity or edibility decisions

This is the important one. 28% of top-1 predictions are wrong. Most errors are same-genus, which usually preserves toxic and edible properties, but a meaningful minority cross families entirely:

Carya glabra (pignut hickory)  ->  Euonymus atropurpureus (burning bush)
Heuchera cylindrica            ->  Bowlesia incana

If your application displays whether a plant is poisonous or safe to eat, an identification from this model alone must not drive that claim. Show a name, and confirm against a larger model or a curated database before making any safety statement. Foraging or medicinal decisions based on this model could cause serious harm.

Error pattern

The model learned genus-level botanical structure well and fails at the final species split. The most frequent confusions:

Xanthium orientale       ->  Xanthium strumarium
Toxicodendron rydbergii  ->  Toxicodendron radicans
Abies sibirica           ->  Abies alba
Ulmus thomasii           ->  Ulmus americana
Oenothera rubricaulis    ->  Oenothera biennis
Fallopia scandens        ->  Fallopia convolvulus

This is the expected failure mode for a 55x parameter reduction. BioCLIP 2.5 uses a ViT-H trained on 200M images precisely because separating visually similar species rewards scale, and an 11.6M parameter student cannot hold that. Design around it with a two-stage system rather than trying to tune it away.

The lookup table has its own ceiling

Some species names sit almost on top of each other in the teacher's text space. Viola tricolor and Viola bicolor score 0.937 cosine similarity as text alone. No image encoder can separate entries that the lookup table cannot distinguish, so these are a property of the approach, not of this student.

Coverage and bias

The 4,271 species are those present in iNaturalist 2021 train_mini under kingdom Plantae, at 50 images each. Coverage reflects where iNaturalist users photograph plants, which skews toward North America and Europe. Species absent from the table cannot be predicted at all, and will instead return whichever listed species is nearest, with no signal that the true answer was missing.

Evaluation caveat

top1_agree measures agreement with the teacher, not correctness against ground-truth labels. Where BioCLIP 2.5 itself is wrong, this metric rewards the student for reproducing that error.

Files

file	what it is
flora_student_fp16.onnx	the model to ship, 23.8 MB
flora_student_fp32.onnx	full precision reference, 47.0 MB
student_weights.pt	PyTorch weights, for further fine-tuning
taxa_table.npy	[4271, 1024] float32 L2-normalized text embeddings
taxa_labels.json	parallel list of species names
eval_report.json	the numbers above, as produced by the eval script

Usage

Input is a [1, 3, 224, 224] float32 tensor with RGB values in 0..1. ImageNet normalization is folded into the graph, so do not apply it yourself. Output is a [1, 1024] L2-normalized embedding.

import json
import numpy as np
import onnxruntime as ort
from PIL import Image

sess = ort.InferenceSession("flora_student_fp16.onnx")
table = np.load("taxa_table.npy")                      # [4271, 1024]
labels = json.load(open("taxa_labels.json"))

img = Image.open("plant.jpg").convert("RGB").resize((224, 224))
x = (np.asarray(img, dtype=np.float32) / 255.0).transpose(2, 0, 1)[None]

emb = sess.run(None, {"image": x})[0][0]               # already unit length
scores = table @ emb                                   # cosine similarity
top5 = np.argsort(-scores)[:5]

for i in top5:
    print(f"{scores[i]:.3f}  {labels[i]['scientific']}")

Because both the embedding and the table rows are unit vectors, the matrix multiply is exactly cosine similarity. On Android, ONNX Runtime Mobile runs the same graph.

License and attribution

This model is released under the MIT License, matching its teacher.

Teacher. BioCLIP 2.5 by the Imageomics Institute, MIT licensed. If you use this work, please cite the BioCLIP 2.5 paper, and consider citing OpenCLIP and the original BioCLIP as that model card requests.

Training data. iNaturalist 2021 competition images, individually licensed by their photographers under various Creative Commons licenses or CC0. Photographers retain copyright. No images are redistributed here, only weights derived from them. Whether model weights constitute a derivative work of their training images is not settled law, and users with strict compliance needs should form their own view.

Built on. OpenCLIP (MIT), timm (Apache 2.0) for the FastViT implementation, and Apple's MobileCLIP work, which established the mobile CLIP distillation recipe this follows.

Citation

@misc{hamilton2026bioclipmobile,
  author = {Hamilton, Nate},
  title  = {BioCLIP 2.5 Mobile: a FastViT student for on-device plant identification},
  year   = {2026},
  url    = {https://github.com/CrazedCoderNate/bioclip-mobile-distill}
}