Transformer Lab

Looks like it’s confirmed Cursor’s new model is based on Kimi! It reinforces a couple of things: - open-source keeps being the greatest competition enabler - another validation for chinese open-source that is now the biggest force shaping the global AI stack - the frontier is no longer just about who trains from scratch, but who adapts, fine-tunes, and productizes fastest (seeing the same thing with OpenClaw for example).

NVIDIA DGX support is now live in Transformer Lab. Got an NVIDIA DGX Spark? Skip the hassle of setting up CUDA 13 and other ML libraries on your machine. Transformer Lab handles environment setup while managing your entire workflow: training/fine-tuning/evals, tracking runs, storing datasets/checkpoints and more. Try it out. Open source, free to use. Feedback welcome. lab.cloud/docs/install/

If you’re setting up an ML research cluster, we wrote this guide for you: The Definitive Guide to Building a Machine Learning Research Cluster From Scratch: • Technical blueprint for single “under-the-desk” GPU server to scaling university-wide cluster for 1,000+ users • Tried and tested configurations for drivers, orchestration, storage, scheduling, and UI with a bias toward modern, simple tooling that is open source and easy to maintain. • Step-by-step install guides (CUDA, ROCm, k3s, Rancher, SLURM/SkyPilot paths) We’ve helped research labs of all sizes build their ML platforms with the goal to create a unified environment to unlock researchers to do their best work. Different budgets, different constraints, but the same questions come up: • How do we evolve from a single workstation into shared compute gracefully? • Selecting an orchestrator / scheduler: SLURM vs. SkyPilot vs. Kubernetes vs. Others? • What storage approach won’t collapse once data + users grow? • How do we avoid building a fragile set of scripts that are hard to maintain? Read the full guide on GitHub (PRs/issues welcome): github.com/transformerlab…

This is amazing to hear -- there is so much opportunity together. Hugging Face has been a terrific steward of open source AI.

6000 commits and we're just getting started!

ML research teams have been stuck with outdated tooling for far too long. Excited to see @TransformerLab launch Transformer Lab for Teams — bringing open, modern infrastructure to real research workflows. Mozilla Ventures is proud to support this team.

@suprgeek Awesome. Chat with our team here if you have any questions: discord.com/invite/transfo…

Trying this right now!

@jameswlepage Excited for you to try it out. Join our Discord to chat with our engineerings if you need help or have questions. discord.com/invite/transfo…

Really excited to try this out

@flappyairplanes Congratulations from @transformerlab ! Love the mission. We hope we get the chance to collaborate one day.

Announcing Flapping Airplanes! We’ve raised $180M from GV, Sequoia, and Index to assemble a new guard in AI: one that imagines a world where models can think at human level without ingesting half the internet.

Microsoft’s FrogMini-14B and FrogBoss-32B now available on Transformer Lab. Fun facts: - FrogMini/FrogBoss are fine-tuned Qwen3 models trained on Claude Sonnet 4 generations - They use multi-turn debugging workflows and complex code reasoning. - They’re great at fixing bugs in Python code Try it out: 1. Install Transformer Lab (free and open source) here: lab.cloud/docs/install/ 2. Select the model from the model registry 3. Run inference, training, evals and generate data

Anyone out there at an ML research lab that works with multiple GPUs? We are launching our biggest new product in a few weeks. We'd love to give you a sneak peak of what we are going to release soon.

The Ultimate Toolkit for Working with LLMs! Transformer Lab lets you train, fine-tune, and chat with any LLM - 100% locally. Enjoy 1-click LLM downloads and a drag-and-drop UI for RAG. 100% open-source.

Can I train a private AI brain on my Mac? What if I could fine-tune a 1B model locally with zero code? Stop paying for generic chat, make AI a contracts expert now I am an online educator, I made a practical, end-to-end walkthrough of fine-tuning an open-source language model locally with Transformer Lab. In this hands-on tutorial, we use Transformer Lab for local finetuning with MLX LoRA on Apple Silicon, from a Hugging Face dataset to a published contracts model, so you master Transformer Lab workflows and local finetuning best practices for open-source LLMs with local finetuning in mind. You will see selecting a small quantized LLaMA-type model, pulling a contracts dataset from Hugging Face, configuring a Jinja2-style prompt template, launching an MLX LoRA training run, monitoring metrics in TensorBoard, evaluating the improved model on a domain-specific prompt about top five contract clauses when buying a house, and publishing results back to Hugging Face. The training job took about 3+ hours on my Mac Studio and can max out RAM and VRAM, so close other apps. Checkpoints save as safetensors, and you can upload LoRA adapters or merge weights for distribution. Why fine-tuning matters, I cover: upskill rapidly, own your data and weights as IP, reduce cost and latency with smaller models, and encode domain guardrails and safety rules into the model. I compare baseline vs fine-tuned outputs to show clear, practical gains, even with a small model. Watch the full walkthrough on YouTube, title Local Finetuning with Transformer Lab: MLX LoRA Walkthrough, and visit givemethemicofficial.com for resoruces and notes. Ask me questions below, I reply. 💻🔧 #transformerlab #localfinetuning #mlxlora Check out this video: youtube.com/watch?v=VyrMYs…

Just shipped text diffusion support in Transformer Lab! Text diffusion models have been getting more attention lately. They generate text by iteratively denoising a noisy or masked sequence, rather than predicting the next token like autoregressive models. Researchers are seeing advantages emerge: more stable training, stronger instruction-following at smaller sizes, fewer token-by-token failure modes, and greater controllability through refinement steps. While still early, diffusion-based language models are becoming a promising alternative for teams exploring new approaches to LLM training and generation. What's included in this release: 🚀 Text Diffusion Server for interactive generation with BERT, Dream, and LLaDA models 🏋️ Text Diffusion Trainer for fine-tuning with masked-language and diffusion-style alignment workflows 📊 Text Diffusion Evaluator for benchmarking with the EleutherAI LM Evaluation Harness We support NVIDIA GPUs for now. AMD and Apple Silicon are coming soon. If you’re exploring diffusion as an alternative to auto-regressive LMs, this gives you a single workspace to do all your experiments. Open source. More info here: lab.cloud/blog/text-diff…

Season 1 of Toronto School of Foundation Modelling is wrapping up soon! 🥹The graduation party will be held on November 27 and will be open to public (more details to come soon). Very grateful for @transformerlab and @aliasaria’s support in making the event possible!!!

We’re heading to @mozillafestival in Barcelona 🇪🇸 (Nov 7–9)! @aliasaria will be speaking on (Re)Wiring Growth While Maintaining Open Source Values → schedule.mozillafestival.org/session/205 Come by Room E1 (Mozilla Festival Social Hub) to chat about the best approaches to train models and run GPU workloads. See you there! #MozFest

Researchers at universities, AI labs, and enterprises have shared with us consistent feedback. As models and datasets grow, the practical work of running experiments becomes harder and more fragmented. Managing hardware, environments, and orchestration layers often takes as much care as designing the models themselves. We’re building an open-source Machine Learning Research Platform: a unified system that abstracts away infrastructure while staying flexible enough for any workflow. It helps teams run distributed training across clusters, manage experiments, and version datasets without the operational overhead. Our users call it “the essential open-source stack for serious ML teams.” Read our vision: 👉 lab.cloud/blog/vision-fo… Need help building ML tooling to overcome a specific challenge? Please reach out. We’re looking for design partners.

I always thought the decline in fundamental AI research funding would happen because AI didn’t generate enough value to be worth the cost. But it seems like it’s happening because it generated too much value. And the race to capture that value is taking priority. Just remembering that a lot of this started in curiosity driven industry research labs.