Tinker
Control every aspect of model training and fine-tuning
426 followers
Control every aspect of model training and fine-tuning
426 followers
Tinker is a flexible API for efficiently fine-tuning open source models with LoRA. It's designed for researchers and developers who want flexibility and full control of their data and algorithms without worrying about infrastructure management.
This is the 3rd launch from Tinker. View more
Inkling 2 Small
Launched this week
Inkling 2 Small is Thinking Machines Lab's new 594B-parameter MoE model with 17B active, built for long-horizon coding, visual tasks, and agents. Available on Tinker for inference and fine-tuning.
Free
Launch Team
Render: Get up to $100K in startup creditsThe fastest way to deploy and scale apps and agents
Promoted
Hi everyone!
Thinking Machines Lab is back with Inkling 2 Small! It's a big jump from the previous generation, especially in coding, visual tasks, and long-running agent work.
In one demo, it spends over four hours training another model to forecast time series. It writes its own evaluator, tries different training approaches, fixes bugs along the way, and picks the best checkpoint.
You can try it and fine-tune it on Tinker now, with full weights coming soon!
And about that name... 594B total parameters, 17B active. I get the MoE math, but 594B is an interesting definition of "Small". Now I'm looking forward to the not-so-small one :)
If it's building a UI for me, does it look at screenshots of the running app to refine it, or only at code and error logs?
The “see and refine what it builds” part is the one I’d actually feel day to day. Most of the time a model hands me SwiftUI code and then I’m the one running it, screenshotting it, and describing what’s off back to the model — the loop is me. If it can look at the rendered result itself and iterate, that changes the workflow completely. The four-hour training demo is impressive, but I’m curious how it holds up on long runs like that — does it stay on the original goal, or does it start drifting and optimizing the wrong thing halfway through?
594B parameters, but only 17B doing the heavy lifting at a time. That's one way to work smarter, not harder.