Transformer Lab

Transformer Lab is a lab building tools intended to move research from a manual, sequential process to a parallelized one, starting with machine learning and AI. Its flagship product, Primus, is presented as an autonomous AI researcher: a user describes a question in a prompt, and the system scopes the question, reviews literature, designs and runs experiments on GPUs it provisions, evaluates results, iterates, and produces a written paper with methods, results, code and citations, alongside artifacts such as trained models and datasets and a record of how each was produced. Projects run in a live workspace where users can watch progress and leave notes that steer subsequent work; runs take hours or days, and Primus scopes experiments to a user-specified budget and states what a larger version would cost. The company says it ran Primus for 30 days, during which it produced more than 30 complete papers spanning LLMs, seismology, physics, crystallography and biology, and that findings Primus helped produce have been cited in published research from labs including DeepMind. Example use cases advertised include fine-tuning open models under hardware constraints, writing fused Triton kernels for MoE routing on H100s, extending an existing arXiv paper, porting models to MLX, and evaluating time-series foundation models on sensor data.

The company also maintains Transformer Lab, an open-source machine learning platform distributed under AGPL-3.0 (transformerlab-app on GitHub), described as an operating system or unified interface for AI research labs. It ships in two editions: an individual edition that runs locally on macOS (Apple Silicon), Linux (NVIDIA or AMD GPU) and Windows via WSL2 with no data leaving the machine, and a Teams edition for labs scaling across GPU clusters, which submits jobs to Slurm clusters or SkyPilot clouds (AWS, GCP, Azure) from one UI and adds centralized experiment tracking, a model registry, artifact management, one-click Jupyter/VSCode/SSH sessions on remote nodes, and auto-recovery from checkpoints and spot-instance preemption. Capabilities span downloading and running open foundation models (Llama 3, DeepSeek, Mistral, Qwen, Phi) across MLX, vLLM, Ollama and HuggingFace Transformers inference engines; format conversion between HuggingFace, GGUF and MLX; chat with multi-turn, batched querying and function calling; full fine-tuning, LoRA/QLoRA, RLHF methods (DPO, ORPO, SIMPO) and reward modeling with trainers for MLX, CUDA and ROCm; YAML-defined hyperparameter sweeps; diffusion and image generation (Stable Diffusion, Flux, ControlNets, IP-Adapters, LoRA adaptor training, WD14 auto-captioning); evaluation via LLM-as-a-judge, the EleutherAI LM Evaluation Harness (MMLU, HellaSwag, GSM8K) and automated red teaming for PII leakage, prompt injection and safety; text-to-speech generation and TTS fine-tuning (Kokoro, Bark); and extensibility through a Python plugin architecture, a Lab SDK for instrumenting existing training scripts, and a CLI.

The GitHub organization hosts 17 repositories, including the main application, an API server, a Python SDK, model/dataset/plugin galleries, examples, an inference repository, a guide to building machine learning research clusters, and per-paper reproducibility packages for experiments (for example a 30B MoE coding model quantized to 24 GB on Apple Silicon, and gravitational-wave detection work).

Founding story

Co-founders Ali Asaria and Tony Salomone, long-time colleagues who previously worked together at Asaria's companies Tulip and Well.ca, began Transformer Lab as a passion project roughly a year before its worldwide launch. They initially set out to do fundamental research, including building a transformer architecture from scratch to compete with large model developers, but repeatedly hit limitations in available tooling; as Asaria described it, they intended to build a research company and instead built the tools to build a research company. A software citation for the project lists Asaria and Salomone as authors with a year of 2023.

Business model

Transformer Lab distributes its research platform as open-source software under AGPL-3.0 in an individual edition that runs on a user's own hardware, and offers a Teams edition positioned as a control plane overlaying existing schedulers and cloud providers for research labs. Primus is delivered as a hosted service at primus.lab.cloud with access granted from a waitlist; the company invites organizations to contact it directly to bring Primus to their teams. Primus experiments run on hardware the service provisions, with users setting a budget for GPU spend.

Traction

The open-source application repository has approximately 5,180-5.2k stars, 547 forks and more than 11,450 commits, with 17 repositories in the GitHub organization. Press coverage at the time of the worldwide launch reported the platform was already in use by some North American research institutes providing feedback, and became broadly publicly available at launch. The company reports Primus produced more than 30 papers in a 30-day run and that resulting findings have been cited in published research from labs including DeepMind. Primus is live with staged onboarding from a waitlist.

Latest developments

Primus was announced as officially live, with users onboarded gradually from a waitlist because of limited GPU capacity, and organizations directed to contact the company for team deployments. Documentation for Primus is published at transformerlab.ai/docs and accounts are created at primus.lab.cloud. Around its worldwide platform launch, the company announced a pre-seed round with participation from Hawktail, Ripple Ventures and Garage Capital, with the amount undisclosed.