Transformer Lab
3 known investors
Transformer Lab is an open source platform for machine learning researchers to request local or cloud compute nodes, submit ML tasks, and track experiments and job artifacts in one place. It supports both individual researchers and teams working across clusters.
Also known as Primus · Transformer Lab · transformerlab
Investors · 3
Also in the syndicate · 1
Company profile
researched Aug 2026Transformer Lab is a Toronto-based company building tools for machine learning and AI research. Its core product is an open-source machine learning platform, described as an operating system or research environment for AI labs, that unifies model training, fine-tuning, evaluation and inference in a single interface and runs locally, on-premises or in the cloud. The individual edition runs natively on macOS (Apple Silicon), Linux and Windows via WSL2 and keeps data on the user's machine; the Teams edition submits jobs to Slurm clusters or SkyPilot clouds (AWS, GCP, Azure) from one UI, provides centralized experiment tracking, a model registry and artifact management, one-click Jupyter, VSCode and SSH sessions on remote nodes, and auto-recovery from checkpoints and spot-instance preemption.
In 2026 the company launched Primus, an autonomous AI research system marketed as an AI researcher available to general users. From a single prompt, Primus scopes a research question, surveys prior literature, designs and runs experiments on GPU compute it provisions, evaluates results, iterates, and produces a written paper with methods, results, code and citations, along with any trained models and datasets the experiments produced and a record of how each was made. Each project has a live workspace where users can follow progress and leave notes that steer subsequent work; projects run over hours or days, and Primus scopes experiments to a stated budget and reports what a larger version would cost. Access is being granted gradually from a waitlist due to limited GPU capacity.
The company frames its mission as accelerating science by turning research from a manual, sequential process into a parallelized utility available to any researcher, starting with machine learning and AI. It states it plans to extend Primus to other scientific domains.
Founding story
Asaria and Salomone set out to do fundamental research, including building a transformer architecture from scratch to compete with large incumbents. In the process they found the available research tooling immature and repeatedly hit limitations, and began building their own tools as they went. As Asaria described it, they intended to build a research company and ended up building the tools to build a research company. The project began as a passion project between the two long-time colleagues.
Business model
Transformer Lab distributes an open-source (AGPL-3.0) research platform in two editions — a local edition for individual researchers and hobbyists running on their own hardware, and a Teams edition that acts as a control plane over existing GPU cluster infrastructure with centralized experiment tracking, model registry and artifact management. Alongside these tools, which help research labs train and evaluate models and manage compute, the company sells access to Primus, its autonomous research system, and is targeting organizations that want to deploy it for their teams. Experiments run on hardware Primus provisions, with projects scoped to a user-specified budget.
The company sells tools that help research labs train and evaluate models and manage compute, and plans to sell Primus to businesses that lack access to machine learning talent, with financial services and pharmaceutical firms cited as potential customers. Pricing is not finalized; the company says it intends to charge a fraction of the cost of employing a human ML researcher with a master's or doctorate, which it puts at up to US$1 million a year. Primus projects consume GPU time scoped to a customer-set budget.
Traction
The open-source repository has approximately 5,200 stars, 547 forks and over 11,400 commits. The platform is in use by research institutes in North America and is publicly available worldwide. Primus produced more than 30 papers in a 30-day internal run across machine learning, seismology, physics, crystallography and biology, one of which was cited by Google DeepMind in a technical report; the company said 30 papers were generated over two months for under $3,000. External users include Rubric Labs. The team consisted of four people as of August 2026.
Latest developments
In August 2026 Transformer Lab launched Primus and began onboarding waitlisted users. The company reported running Primus for 30 days to produce more than 30 papers spanning LLMs, seismology, physics, crystallography and biology, and said Google DeepMind cited a Primus-produced paper in a technical report on one of its new models. Rubric Labs, a Toronto software development agency, used Primus to test ideas for compressing data fed into AI model context. The company said its commercial model was still being worked out, that it intends to price well below the cost of employing a graduate-level ML researcher, and that it plans to extend Primus to scientific fields beyond machine learning.
▸Full profile — market position, technology, go-to-market, geography, history, risks & controversies
Market position
Transformer Lab is a four-person, pre-seed startup competing in the emerging autonomous-science category against much larger and better-funded entrants, including projects at major technology firms and startups founded by prominent researchers such as Jeff Clune and David Silver that have raised hundreds of millions of dollars, as well as Discovery Loop, co-founded by Google's Jeff Dean and co-led by Radical Ventures. The company claims to be more advanced than better-known competitors in areas such as literature grounding and hallucination control, and treats their entry as validation of the category. Its open-source research platform has drawn roughly 5,200 GitHub stars and 547 forks.
The company positions its platform as a unified, modern replacement for fragmented and dated AI research tooling, emphasizing open source licensing, local/privacy-first operation for individuals, and an overlay architecture for teams that augments rather than replaces existing schedulers. For Primus, the company claims advantages in supplying relevant literature — it maintains its own library of research papers weighted by quality rather than relying on general web search — and in limiting hallucination by requiring Primus to produce real experimental data supporting its results and to address critiques from built-in AI reviewer agents. It also frames low cost per paper (under $3,000 for 30 papers, per its CEO) and accessibility to non-specialists as differentiators, arguing that idea quality rather than researcher credentials should determine research output.
Technology
The platform supports downloading and running open foundation models (Llama 3, DeepSeek, Mistral, Qwen, Phi and others) across inference engines including MLX, vLLM, Ollama and HuggingFace Transformers, with conversion between HuggingFace, GGUF and MLX formats. Training features include full fine-tuning, LoRA/QLoRA, RLHF methods (DPO, ORPO, SIMPO), reward modeling, hardware-specific trainers for Apple Silicon (MLX), NVIDIA (CUDA) and AMD (ROCm), and YAML-defined hyperparameter sweeps. It also covers diffusion image generation (Stable Diffusion, Flux, ControlNets, IP-Adapters, LoRA training, WD14 auto-captioning), audio/TTS generation and fine-tuning (Kokoro, Bark), and evaluation via LLM-as-a-judge scoring, the EleutherAI LM Evaluation Harness (MMLU, HellaSwag, GSM8K) and automated red teaming for PII leakage, prompt injection and safety. Extensibility comes from a Python plugin system, a Lab SDK for instrumenting existing training scripts, and a CLI. Orchestration integrates Slurm and SkyPilot. Primus is built on multiple AI agents that read literature, generate hypotheses, write and run code and experiments, and review outputs, drawing on a company-curated corpus of research papers weighted by quality.
Go-to-market
Distribution is led by open source: the platform is published on GitHub under AGPL-3.0 with an install script, CLI, Python plugin architecture and a Lab SDK, plus a Discord community and public documentation. Adoption began with demos to selected users in Canada and the US and early deployments at North American research institutes before a public worldwide launch. Primus is being rolled out through a waitlist with direct contact for organizations, and research papers produced by Primus are published on the company's research page as demonstration of capability. The company also recruits through its Discord.
Machine learning researchers and hobbyists working on single machines; research labs and institutes scaling across GPU clusters; and, for Primus, organizations without in-house ML talent, including financial services and pharmaceutical firms developing models, plus individual users and software agencies who want to test research ideas without owning GPUs.
Geography
Headquartered in Toronto, Ontario, Canada; the GitHub repository notes the software is built in Canada. Investors span Canada and the United States, the platform is available worldwide, and early adopters included research institutes in North America.
History
Ali Asaria and Tony Salomone, former colleagues at Asaria's retail companies Well.ca and Tulip, co-founded Transformer Lab in September 2024. The GitHub repository's citation metadata dates the project to 2023. The pair built early prototypes and demonstrated them to a small group of users in Canada and the US, then raised a pre-seed round from Hawktail, Ripple Ventures, Garage Capital and Mozilla's venture arm, which the company said funded scaling to a global, enterprise-ready launch. The platform launched worldwide and became broadly available to the public, after early use by North American research institutes. In August 2026 the company launched Primus, an autonomous AI research system, and stated its intention to extend Primus beyond machine learning into other scientific fields.
Risks & controversies
The company competes against far better-capitalized entrants pursuing recursive self-improvement and automated science, and its own pre-seed funding amount was not disclosed. Primus availability is constrained by limited GPU capacity, requiring waitlist-based onboarding, and heavier projects consume substantial GPU spend. The commercial model was still being determined as of August 2026. More broadly, the company notes that increasing academic use of generative AI has produced many documented cases of fabricated citations, a failure mode it says Primus is designed to avoid by requiring supporting data and internal reviewer critique.
Compiled by commissioned research from 8 cited public sources — announcements, filings, and press listed under research sources below.
Key figures
latest reportedCompany-reported or press-reported figures, each dated to when it was claimed — not independently audited.
Competitors · 2
by search overlap
Companies competing with Transformer Lab for the same Google search keywords, organic and paid, via search-intersection analysis.
Timeline · 5
launches, deals, and filingsGoogle DeepMind cited a paper written by Primus in a technical report about one of its new models.
Rubric Labs, a Toronto-based software development agency, used Primus to explore methods of condensing data fed into AI models' working memory.
Transformer Lab launched Primus, a system whose AI agents scan scientific literature, form a hypothesis, run experiments on provisioned GPU compute and write an academic-style paper from a single prompt. Users are being onboarded gradually from a waitlist due to limited GPU capacity.
Transformer Lab announced a pre-seed fundraising round of undisclosed amount with participation from Hawktail, Ripple Ventures and Garage Capital, coinciding with the worldwide launch of its platform. Mozilla's venture arm is also listed among backers.
The AI research platform, positioned as a unified modern operating system for AI research labs, became broadly available to the public, having previously been used by research institutes in North America.
Dated company events from announcements, filings, and press; legal rows summarize public dockets and regulator releases.
In the news
▸Research sources · 8
primary sources listed
- Transformer Lablab.cloud · web
8 public sources were cited for this profile; the first-party ones are listed here.
Frequently asked questions
- What does Transformer Lab do?
- Toronto startup building an open-source AI research platform and Primus, an autonomous AI researcher that runs experiments and writes papers.
- Who are Transformer Lab's investors?
- Transformer Lab's investors include HAWKTAIL, Ripple Ventures.