·
Researchers & Thinkers
Tim Dettmers
Assistant Professor (MLD & CSD) |Carnegie Mellon University / Ai2
Creator of bitsandbytes and QLoRA. Pioneering k-bit quantization for accessible LLM inference and fine-tuning. Leading open-source coding agents (SERA) at Ai2.
3papers
GitHubTimDettmersResearch Papers
3 papersBiography
Deep Research
Tim Dettmers is an Assistant Professor at Carnegie Mellon University (Machine Learning and Computer Science departments) and a Research Scientist at the Allen Institute for Artificial Intelligence (Ai2). He earned his PhD from the University of Washington under Luke Zettlemoyer. Dettmers is the creator and maintainer of bitsandbytes (8,000+ stars, 2.2 million monthly installs), the foundational open-source library for k-bit quantization in PyTorch that powers efficient LLM inference and fine-tuning across the ecosystem. He is the lead author of QLoRA, an efficient fine-tuning method that enables training a 65B-parameter model on a single 48GB GPU while preserving full 16-bit performance, and LLM.int8(), which brought 8-bit matrix multiplication to transformers at scale. His current research focuses on open-source coding agents (SERA) and making foundation models accessible on consumer hardware. Before his PhD he worked three years in factory automation and studied psychology. He describes himself as dyslexic with bottom-5% working memory, which he credits for driving him toward simpler, more elegant solutions.
Quantization (4-bit, 8-bit, NormalFloat)Efficient LLM Fine-Tuning (QLoRA, LoRA)bitsandbytes LibraryOpen-Source Coding Agents (SERA)LLM Inference OptimizationSparse LearningKnowledge Graph Embeddings (ConvE)AI Accessibility & DemocratizationHardware Scaling LimitsOn-Device Mixture of Experts
Timeline
13 Research13 total
2026
2025
2024
2023
2022
2021
2021-01Research
Received NeurIPS 2021 Best Reviewer Award2018
Key Contributions
bitsandbytes
Foundational open-source library for k-bit quantization in PyTorch. Enables accessible LLM inference and training via 4-bit and 8-bit optimizers. 8,000+ stars, 2.2 million monthly installs, received Google Open Source and PyTorch Foundation awards.
QLoRA
Efficient fine-tuning method using 4-bit NormalFloat quantization and double quantization that enables training a 65B-parameter model on a single 48GB GPU. Guanaco models reached 99.3% of ChatGPT performance on Vicuna benchmark. 10,800+ stars. NeurIPS 2023.
LLM.int8()
8-bit matrix multiplication method for transformers at scale, enabling efficient inference for billion-parameter models without significant quality degradation. A key building block in the quantization ecosystem.
SERA (Soft-Verified Efficient Repository Agents)
Open-source coding agent from Ai2 that solves 54.2% of SWE-Bench Verified. Built with 32 GPUs and 5 researchers, reproducible for ~$400, 26x more efficient than RL approaches.
ConvE
Convolutional 2D Knowledge Graph Embeddings -- a pioneering approach for link prediction in knowledge graphs using 2D convolutions over embedding matrices. 692 stars.
Sparse Momentum
Sparse learning library implementing sparse momentum for training sparse neural networks. 385 stars.
Notable Quotes
“
The thinking around AGI and superintelligence is not just optimistic, but fundamentally flawed.
“
We have maybe one, maybe two more years of scaling left before further improvements become physically infeasible.
“
Sometimes constraints force you to find simpler solutions -- and sometimes those solutions turn out to be better.
“
Because our method is cheap, it opens coding agent research to everyone. You do not need large teams or thousands of dollars.
“
Switching fields from one that you are well established in to something completely new is probably one of the hardest things you can do in research.
“
Open source can be competitive and might actually overtake closed source APIs.
15 sources(click to expand)
About Me -- Tim Dettmers (personal website)Tim Dettmers -- CMU Faculty ProfileTimDettmers on GitHub (59 repos, 1.4K followers)bitsandbytes-foundation/bitsandbytes on GitHub (8K+ stars)QLoRA: Efficient Finetuning of Quantized LLMs (arXiv 2305.14314)QLoRA NeurIPS 2023 proceedingsSERA: Soft-Verified Efficient Repository Agents (arXiv 2601.20789)Interviewing Tim Dettmers on open-source AI (Interconnects)Ai2's Tim Dettmers: AGI is a fantasy (The Register)Dettmers Receives Google ML and Systems Junior Faculty Award (CMU)Tim Dettmers & Aviral Kumar Named AI2050 Early Career Fellows (CMU)My Journey Towards Coding Agents: Building SERA (blog)Tim Dettmers on Google ScholarMaking LLMs accessible with bitsandbytes and QLoRA (Hugging Face blog)Open Coding Agents (Ai2 blog)
Research generated March 19, 2026
Researchers & Thinkers/Tim Dettmers
All Profiles