Research and results

We publish what we measure, how we measured it and where the method stops. The reports are kept with the code they describe.

Software extension of GPU memory

V²RAM manages the residency of tensor blocks one level below the framework executor, inside the buffer object. Blocks move between the card, pinned host memory and a memory-mapped store under a budget, guided only by the sequence of accesses that the run reveals. No network, layer or training loop changes.

Cost against keeping everything on the card
ConfigurationBudgetResidentWith V²RAMRatio
Inference, working set of 200 MiB128 MiB13.2 ms35.2 ms2.7×
Training, working set of 1,186 MiB128 MiB256.6 ms353.7 ms1.4×

Milliseconds per step, median of three runs. The budget is set below the working set on purpose.

Capacity of a card of 12,288 MiB with a budget of 8,192 MiB
BatchWorking setPeak on the cardOutcome
Without V²RAM6,028 MiB–Did not start: host out of memory
16,028 MiB10,748 MiBRan
27,709 MiB10,170 MiBRan
410,984 MiB9,568 MiBRan
817,506 MiB9,178 MiBRan
1630,600 MiB (estimated)9,092 MiBStopped by the system

A model of 270 million parameters with grouped-query attention and four experts per layer.

Detection and replacement of personal data

Four devices share the detection: scanners that verify check digits, word tables in five languages, a structural SVM tagger and a byte-level transformer of 116,791 parameters. Each value is replaced by a tag that carries its encryption under FF1 and AES-256.

Version 1.0.11 reaches an F1 of 0.946 on the test set, 0.919 with six times the noise, 0.953 on whole documents and 0.951 on places abroad; the tables give the detailed measures of version 1.0.5. On hand-written texts in the five languages, the F1 lies between 0.886 and 0.907. On sixteen files written for the test in the five languages (Word, PowerPoint, PDF, CSV, source code, images), version 1.0.11 leaves readable 2 of the 1,223 values to protect, against 109 for the previous version. On a private sample of real documents, precision is 0.859 and recall 0.912: about one value in eleven stays readable, which is why a protected document should be read again before it is shared. The detector reads about 41,000 characters per second on two cores.

Whole detector, exact spans, every category (Pseudalis 1.0.5)
Test setPrecisionRecallF1
Test, five languages0.9550.9350.945
Values never seen in training0.9510.9350.943
Six times the noise0.9410.8970.918
Whole documents0.9610.9430.952
Places abroad only0.9600.9390.949
By languagePrecisionRecallF1
French0.9750.9650.970
English0.9710.9640.967
Italian0.9500.9260.938
German0.9490.9260.937
Spanish0.9280.8980.913
Contribution of each component (Pseudalis 1.0.5)
SystemF1 on the test set
Rules alone0.840
SVM tagger alone0.749
Byte-level transformer alone0.903
Whole detector0.945

Inference on quantized weights

The engines of Albireo keep the weights in the memory mapping of the GGUF file and multiply them in their quantized form. Integer kernels, a cache of keys and values in 8 bits, drafted tokens and layers computed on the card are each measured against a reference.

  • Keys and values in 8 bits give the same next token as half precision on the 210 positions of a text, with a mean divergence of 6.5e-5.
  • Drafted tokens rewrite a function at 25.5 tokens per second against 19.7 on two cores, and keep the distribution of the sampling.
  • Four answers written together take 32 to 40% less time per step than four separate ones.
  • On the 24 threads of an Intel Core Ultra 9 275HX, the processor engine reads a prompt at 216 tokens per second and writes at 19.5 with Qwen3-VL 4B; with Qwen3 0.6B it reads at more than 1,300 tokens per second and writes at more than 80.
Generation, Qwen3-VL 4B in Q4_K_M
EngineTokens per second
NVIDIA CUDA, GeForce RTX 5070 Ti Laptop GPU56.8
Processor, 24 threads23.6
Reading a prompt of 512 tokens, Qwen3 0.6B, two cores
KernelsIntegersFloats
AVX-VNNI230 tokens/s126 tokens/s
AVX2154 tokens/s125 tokens/s

Routing of operations to the NPU and the integrated GPU

Built with the OpenVINO runtime, the library offers each tensor operation to the inference devices of the processor. The host and each device are timed on a few calls, and the fastest keeps that shape. A machine without these devices runs exactly as before.

Training step of a small transformer, Intel Core Ultra 9 275HX
ConfigurationTime per step
Host alone, with Intel oneMKL97.4 ms
Automatic routing63.3 ms

The routing hands softmax, activations, cross entropy and convolutions to the integrated GPU, which computes them 2.2 to 6.9 times faster than the host. The NPU stays a candidate for every operation: at the sizes of this benchmark its fixed cost per call still outweighs its arithmetic, and the routing selects it as soon as a measurement finds it the fastest.

Prediction as compression

A trained network is a probability model of what comes next, and an arithmetic coder turns it into a lossless code. One example builds the whole chain: a tokenizer and a decoder trained in Dlib, weights quantized and compiled into the program, and arithmetic fixed operation by operation so that a decoder on another machine reproduces every probability.

enwik9, the gigabyte of English Wikipedia text of the Large Text Compression Benchmark
Archive117,169,624 bytes
Bits per byte0.937
Decompressor, tokenizer and network included7,063,888 bytes
Time per direction, one coreAbout 25 hours
Compression with the kernels chosen by the processor24 h 48 min instead of 30 h 18 min

Hierarchical reasoning and adaptive computation

Two example programs train hierarchical reasoning models on Sudoku-Extreme and on ARC puzzles. They bring to the library a halting head trained jointly with the network, the Adam-atan2 solver, a stablemax loss and an average of the weights, each available to the other models of the repository.

Images: self-supervision, colorization and synthesis

Our earlier work on images produced methods and models that remain published.

WorkResult
BAREL, self-supervised learning followed by supervised metric learning96.1% on the LFW benchmark for face recognition, with a training set nearly 8 times smaller than that of the first ResNet model of Dlib
Barlow Twins on CIFAR-10 with a ResNet-5088.25% test accuracy after self-supervised training
Colorify, high-resolution colorizationA network of 223 layers and about 41.7 million parameters trained on about 900,000 images
FFgan, synthetic facesColor faces of 162 pixels from a DCGAN trained on over 150,000 faces

Contributions to Dlib

Dlib is one of the reference open-source C++ libraries for machine learning and computer vision. Before publishing our own extensions, we contributed code and trained models to the library itself, first for faces (recognition, gender, age, synthesis), then for transformers.

In the library

Fourteen contributions were merged into Dlib between August 2024 and October 2025.

  • Layers for transformers: RMS normalization, embeddings, linear, lower triangular mask, transposition, reshaping, plane-wise matrix product
  • An adaptive computation time layer and a dropout layer with a rate set at compile time
  • A byte-pair encoding tokenizer
  • Example programs: a transformer network, then one with rotary position embeddings and a mixture of experts
  • A loader for the ARC-AGI dataset
  • A fix to the reorg layer, on processor and on CUDA
In the official model repository
ModelTaskResult
face_recognition_densenet_model_v1Face recognition, DenseNet trained with the BAREL method96.1 % on LFW, with a training set nearly 8 times smaller than that of the first ResNet model of Dlib
dnn_gender_classifier_v1Gender classification from a faceAbout 97.3 % on the LFW evaluation, trained on about 200,000 faces
dnn_age_predictor_v1Age estimation from a face, ResNet-10Trained on about 110,000 labelled images
dcgan_162x162_synth_facesSynthetic faces, DCGANColour faces of 162 pixels
highres_colorifyAutomatic colorization, ResNet with a U-Net structure223 layers and about 41.7 million parameters, usable with an 8 GB graphics card
res50_self_supervised_cifar_10Self-supervised learning, Barlow Twins, ResNet-5088.25 % test accuracy on CIFAR-10
resnet34_stable_imagenet_1kImage classification, ResNet-3499.98 % top-1 on the test split of Stable ImageNet-1K
vit-s-16_stable_imagenet_1kImage classification, Vision Transformer99.76 % top-1 on the test split of Stable ImageNet-1K

The gender and age models are published by Cydral Technology under the Creative Commons Zero 1.0 licence.

Publications

Buffer-Level Residency Management for GPU Memory Extension
Aldric Pierrain, Cydral Technology. Research report, September 2026
Technical application architecture of Pseudalis Community
Aldric Pierrain, Cydral Technology. Research report, November 2026
Onward to the Hutter Prize: a token-level transformer compresses enwik9 to 117 MB on one CPU core
Aldric Pierrain, Cydral Technology. Research report, November 2026
BAREL: Backpropagation And Reinforcement-based Environment for Learning
Aldric Pierrain, Cydral Technology. Publication, April 2023
A Small Language Model Stack, Built in C++ on One Machine
Aldric Pierrain, Cydral Technology. Article

Measurement method

Each figure names its hardware, its model and its protocol. Results are cross-checked against reference implementations before they are stated, and the limits of a method are written in the same place as its results.

This work pursues the objectives of P16, the France 2030 programme led by Inria for sovereign AI libraries.