Mahyar Najibi

Co-founder & Chief Scientific Officer

ElastixAI · Seattle, WA

I am a Co-founder of ElastixAI, where we are building a new foundation for AI inference. Our platform is built on reconfigurable hardware, letting us co-design machine learning, systems software, and the hardware itself to stay future-proof as models and optimization techniques evolve, which steadily reduces both cost per token and energy per token.

Before founding ElastixAI, I was a Senior AI/ML Manager at Apple, where I led teams advancing language and vision foundation models. Earlier, I was a Lead Scientist at Waymo Research, building perception and prediction systems that powered autonomous driving, and I worked on 3D perception at Google Research before that. I earned my Ph.D. in Computer Science from the University of Maryland.

Prospective Candidates: We are hiring machine learning scientists, machine learning engineers, and AI infrastructure engineers at ElastixAI. If that's you, don't hesitate to reach out, and apply through our careers page.

News

2026We are hiring! ElastixAI is looking for machine learning scientists, machine learning engineers, and AI infrastructure engineers, see our careers page.
2026The ElastixAI blog is live — check it out for our latest posts.
2025We founded ElastixAI, building a new foundation for AI inference.
2025The Apple Intelligence Foundation Models tech report is out.
2025QuantSpec accepted to ICML 2025.
2025SeedLM accepted to ICLR 2025.
2024Superposition Prompting accepted to ICML 2024.
2024Check out our new work on efficient LLM inference: Speculative Streaming.
2024Several new works on efficient and multimodal models: OpenELM, Apple Intelligence Foundation Models, LazyLLM, CatLIP, and CtrlSynth.
2023One paper accepted to ICCV 2023.
2023Two papers accepted to CVPR 2023.
2022Two papers accepted to ECCV 2022.
2021One paper accepted to NeurIPS 2021.
2021I received the Outstanding Reviewer Award at NeurIPS.
Feb 2021One paper accepted to CVPR 2021.
Feb 2021Our manuscript on multi-scale detection is accepted by PAMI.
Feb 2020One paper accepted to CVPR 2020.
Nov 2019Two papers accepted to AAAI 2020.
Sep 2019One paper accepted to NeurIPS 2019.
Jul 2019One paper accepted to ICCV 2019.
May 2019Two papers accepted to BMVC 2019.
May 2019One paper accepted to CVPR 2019.
Dec 2018Check out our new work on fast multi-scale inference: AutoFocus.
Sep 2018Two papers accepted at NIPS 2018.
Jun 2018The code for the SNIPER detector is released: github.com/MahyarNajibi/SNIPER.
May 2018Check out our new work on detection: SNIPER: Efficient Multi-Scale Training.
May 2018I received the Outstanding Reviewer Award at CVPR 2018.

Publications

arXiv '26

LEAP: Learnable End-to-End Adaptive Pruning of Large Language Models

M. Mozaffari, Y. Hourri, M. Rastegari, Mahyar Najibi

arXiv 2026

Mahyar Najibi

News

Publications

LEAP: Learnable End-to-End Adaptive Pruning of Large Language Models

M2R2: Mixture of Multi-Rate Residuals for Efficient Transformer Inference

QuantSpec: Self-Speculative Decoding with Hierarchical Quantized KV Cache

From Dense to Dynamic: Token-Difficulty Driven MoEfication of Pre-trained LLMs

SeedLM: Compressing LLM Weights into Seeds of Pseudo-Random Generators

Speculative Streaming: Fast LLM Inference without Auxiliary Models

Efficient Vision-Language Models by Summarizing Visual Tokens into Compact Registers

CtrlSynth: Controllable Image-Text Synthesis for Data-Efficient Multimodal Learning

Apple Intelligence Foundation Language Models

LazyLLM: Dynamic Token Pruning for Efficient Long Context LLM Inference

CatLIP: CLIP-level Visual Recognition Accuracy with 2.7× Faster Pre-training on Web-scale Image-Text Data

OpenELM: An Efficient Language Model Family with Open Training and Inference Framework

Superposition Prompting: Improving and Accelerating Retrieval-Augmented Generation

Unsupervised 3D Perception with 2D Vision-Language Distillation for Autonomous Driving

3D Human Keypoints Estimation from Point Clouds in the Wild without Human Labels

GINA-3D: Learning to Generate Implicit Neural Assets in the Wild

Motion Inspired Unsupervised Perception and Prediction in Autonomous Driving

Improving the Intra-class Long-tail in 3D Detection via Rare Example Mining

Revisiting 3D Object Detection from an Egocentric Perspective

Offboard 3D Object Detection from Point Cloud Sequences

Scale Normalized Image Pyramids with AutoFocus for Object Detection

DOPS: Learning to Detect 3D Objects and Predict their 3D Shapes

Universal Adversarial Training

Generate, Segment, and Refine: Towards Generic Manipulation Segmentation

Adversarial Training for Free!

AutoFocus: Efficient Multi-Scale Inference

FA-RPN: Floating Region Proposals for Face Detection

Soft Sampling for Robust Object Detection

Batch-wise Logit-Similarity: Generalizing Logit-Squeezing and Label-Smoothing

An Analysis of Pre-Training on Object Detection

SNIPER: Efficient Multi-Scale Training

Poison Frogs! Targeted Clean-Label Poisoning Attacks on Neural Networks

Towards the Success Rate of One: Real-time Unconstrained Salient Object Detection

Face-MagNet: Magnifying Feature Maps to Detect Small Faces

SSH: Single Stage Headless Face Detector

G-CNN: an Iterative Grid Based Object Detector

Supervised Incremental Hashing

On Large-Scale Retrieval: Binary or n-ary Coding?

Patents

Pseudo Random Projection for Machine Learning Compression

Saliency-based Object Counting and Localization

Deep Patch Feature Prediction for Image Inpainting