·
Builders & Technical Leaders
Younes Belkada
Senior AI Engineer |Technology Innovation Institute (TII)
Core developer of PEFT, TRL, and bitsandbytes integration at Hugging Face. Co-author of LLM.int8() and Falcon-H1.
7papers
GitHubyounesbelkadaResearch Papers
7 papersarXiv
LLM.int8(): 8-bit Matrix Multiplication for Transformers at Scale
2208.07339·2022-08-15
arXiv
Petals: Collaborative Inference and Fine-tuning of Large Models
2209.01188·2022-09-02
arXiv
BLOOM: A 176B-Parameter Open-Access Multilingual Language Model
2211.05100·2022-11-09
arXiv
Zephyr: Direct Distillation of LM Alignment
2310.16944·2023-10-25
arXiv
StarCoder 2 and The Stack v2: The Next Generation
2402.19173·2024-02-29
arXiv
Falcon Mamba: The First Competitive Attention-free 7B Language Model
2410.05355·2024-10-07
arXiv
Falcon-H1: A Family of Hybrid-Head Language Models Redefining Efficiency and Performance
2507.22448·2025-07-30
Biography
Deep Research
Younes Belkada is a Senior AI Engineer at the Technology Innovation Institute (TII) in Abu Dhabi, where he works on pre-training, evaluation, and tooling for the Falcon family of large language models. Previously he spent three years (2021-2024) as a Machine Learning Engineer on the Hugging Face Open Source team, where he became a core developer of PEFT (Parameter-Efficient Fine-Tuning), TRL (Transformers Reinforcement Learning), and the bitsandbytes quantization integration in Transformers. He co-authored the landmark LLM.int8() paper with Tim Dettmers, pioneered the native Flash Attention 2 integration in Hugging Face Transformers, and wrote the widely-used 4-bit QLoRA integration blog post. He holds an MSc in Mathematics, Vision, and Learning (MVA) from ENS Paris-Saclay and studied Applied Mathematics and Computer Science at Polytech Sorbonne (Sorbonne Universite), with an exchange semester in Data Science at EPFL. He has co-authored 10+ papers spanning BLOOM, StarCoder 2, Zephyr, Petals, Falcon Mamba, and Falcon-H1, and co-instructed the DeepLearning.AI course 'Open Source Models with Hugging Face'.
Parameter-Efficient Fine-Tuning (PEFT / LoRA / QLoRA)Model Quantization (bitsandbytes, LLM.int8, 4-bit NF4)Transformers Reinforcement Learning (TRL / RLHF / DPO)Flash Attention IntegrationFalcon LLM Family (Falcon Mamba, Falcon-H1, Falcon 3)Hybrid Transformer-SSM ArchitecturesLarge-Scale Pre-training (BLOOM 176B, StarCoder 2)Distributed Inference (Petals)LLM Alignment (Zephyr, Direct Distillation)Open-Source ML Tooling & Developer Education
Timeline
15 Research15 total
2026
2025
2024
2023
2023-05Research
Co-authored HF blog post 'Making LLMs even more accessible with bitsandbytes, 4-bit quantization and QLoRA' (176 upvotes)2023-09Research
Integrated native Flash Attention 2 support into Hugging Face Transformers, acknowledged by Tri DaoKey Contributions
PEFT (Parameter-Efficient Fine-Tuning)
Core developer of Hugging Face's PEFT library (20.8k stars) enabling LoRA, QLoRA, and other parameter-efficient methods for fine-tuning large models on consumer hardware
bitsandbytes / QLoRA Integration in Transformers
Built the native 4-bit and 8-bit quantization integration in Hugging Face Transformers via bitsandbytes, making it possible to load and fine-tune 65B-parameter models on a single 48GB GPU
TRL (Transformers Reinforcement Learning)
Core contributor to the TRL library (17.7k stars) for training transformer language models with RLHF, DPO, and PPO
LLM.int8() — 8-bit Matrix Multiplication at Scale
Co-authored the LLM.int8() paper (NeurIPS 2022) introducing mixed-precision decomposition that halves inference memory without accuracy loss, enabling 175B models on consumer GPUs
Flash Attention 2 Integration in Transformers
Led the native integration of Flash Attention 2 into Hugging Face Transformers, enabling faster and more memory-efficient training and inference across 30+ model architectures
Falcon-H1 Hybrid Language Models
Co-authored the Falcon-H1 family of hybrid Transformer-SSM models at TII, spanning 0.5B to 34B parameters with state-of-the-art efficiency
BLOOM 176B
Contributed to the BigScience Workshop's BLOOM, a 176B-parameter open-access multilingual language model trained collaboratively by 1000+ researchers
Open Source Models with Hugging Face (DeepLearning.AI)
Co-instructed the DeepLearning.AI short course teaching NLP, audio, image, and multimodal tasks using open-source Hugging Face models
Notable Quotes
“
Use PEFT (Parameter-Efficient Fine-Tuning)! In the Hugging Face ecosystem, you can now fine-tune large language models with a fraction of the memory using LoRA and QLoRA.
“
GPTQ quantized models can be now loaded out of the box in transformers, making large model inference accessible to everyone.
“
The Falcon has landed in the Hugging Face ecosystem.
12 sources(click to expand)
younesbelkada on GitHub (91 repos, 861 followers)ybelkada on Hugging Face (144+ models, 25 spaces, 11 papers)Younes Belkada on LinkedInLLM.int8(): 8-bit Matrix Multiplication for Transformers at Scale (NeurIPS 2022)Making LLMs even more accessible with bitsandbytes, 4-bit quantization and QLoRA (HF Blog)Falcon-H1: A Family of Hybrid-Head Language Models (arXiv)NeurIPS 2025 E2LM Competition (arXiv)Open Source Models with Hugging Face (DeepLearning.AI)Younes Belkada on CatalyzeX (publications list)Younes Belkada's personal blog (about page)Tri Dao acknowledging Flash Attention 2 HF integration (X/Twitter)Younes Belkada on ResearchGate (MVA, ENS Paris-Saclay)
Research generated March 19, 2026
Builders & Technical Leaders/Younes Belkada
All Profiles