EULOGIK EDGE MODELFeatured

pico-type

v02 (ArXiv:2608.14658) · 1.5 Million
Target HardwareAny CPU / ARM64 / Microcontroller
Latency / Speed<6ms (CPU AVX2) · <12ms (Raspberry Pi)
RAM / Memory8 MB RAM
Fertility / Size200 KB ONNX Binary
  • Tokenizer-free byte architecture: immune to OOV words
  • 7 simultaneous classification heads in a single 200KB file
  • Classifies 62 programming languages & 90 MIME types
  • Published research paper on ArXiv (2608.14658)
Download Open Weights on HuggingFace ↗
INDUSTRY BASELINEOpen Weights

FastText

lid.176.bin · N/A (Embedding Matrix)
Target Hardwarex86 / Server CPU
Latency / Speed2–4ms (CPU)
RAM / Memory150–250 MB RAM
Fertility / Size126 MB to 200 MB Binary
  • Extremely fast subword embeddings
  • 1,000× larger binary size (126MB vs 200KB)
  • Requires external tokenization and text preprocessing
  • Single task only (language ID or text category)
METRIC SPECIFICATION MATRIX
METRICPICO-TYPEFASTTEXT
Binary Export Size200 KB (ONNX) 126 MB (.bin)
Tokenizer DependencyNone (Byte-level) Required
Multi-Head Multi-Task7 Heads (Lang + MIME + Code) Single Head
RAM Footprint8 MB 180 MB
Inference Latency (CPU)<6 ms 3 ms
REPRODUCIBLE CODE SNIPPET

Execute this benchmark locally

import onnxruntime as ort
import numpy as np

# Load 200KB ONNX model
session = ort.InferenceSession("pico_type_v02.onnx")
raw_bytes = np.frombuffer("def calculate_tax(amount): return amount * 0.18".encode("utf-8"), dtype=np.uint8)
inputs = {"input_bytes": raw_bytes[np.newaxis, :512]}
outputs = session.run(None, inputs)
print("Detected Language:", outputs[0]) # Output: Python in <6ms

Interested in deploying pico-type?

We assist enterprises with custom fine-tuning, hardware sizing, and on-premise air-gapped installation.