Announcement: MLX Fork

[gajsanders]

Hi all,

I’ve completed a Mac MLX-compatible port of TinyRecursiveModels. The new fork supports Apple Silicon, is optimized for MLX, and includes test-driven workflows, CLI tools, and reproducible datasets.
The MLX port includes reproducible mini-datasets for testing and experimentation, but full support for larger external datasets (e.g., maze and sudoku) is on the roadmap.

You can find the project here: https://github.com/gajsanders/TinyRecursiveModels-MLX

Feedback and suggestions welcome!

[olivkoch]

Nice work! Just curious: my experience is that TRM requires a big GPU to train a model on a decent problem (Sudoku Extreme, Maze Hard, etc.) What is the purpose of your fork in that context? Is that to train smaller TRM models on a Mac laptop?

[gajsanders]

Thanks for the kind words! You're spot on—original TRM training on tougher benchmarks like Sudoku Extreme or Maze Hard does demand serious GPU power, like multiple H100s. My fork tweaks the setup for faster iteration on smaller puzzles (e.g., 4x4/6x6 Sudoku) with streamlined configs via Hydra/UV/Lightning and on-the-fly data generation. This lets me prototype and train compact TRM variants right on a MacBook's unified memory without cloud resources—ideal for local experimentation before scaling up. Curious about your setup for those bigger problems?

…

On Thu, Dec 11, 2025 at 9:47 AM Olivier K ***@***.***> wrote: *olivkoch* left a comment (SamsungSAILMontreal/TinyRecursiveModels#50) <#50 (comment)> Nice work! Just curious: my experience is that TRM requires a big GPU to train a model on a decent problem (Sudoku Extreme, Maze Hard, etc.) What is the purpose of your fork in that context? Is that to train smaller TRM models on a Mac laptop? — Reply to this email directly, view it on GitHub <#50 (comment)>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/AT7SXSNT5R2SUEJDLAMR3AD4BEVRHAVCNFSM6AAAAACMP5TKW6VHI2DSMVQWIX3LMV43OSLTON2WKQ3PNVWWK3TUHMZTMNBQHA3DENJQHA> . You are receiving this because you authored the thread.Message ID: ***@***.***>

[olivkoch]

Sudoku Extreme does not require multiple H100s. I'm using a single H100 but an A100 might be enough.

I'm not sure what you intend to do with on-the-fly data generation. In my experience, it makes your experimentation a lot slower. Generating sudokus is surprisingly compute-intensive. Unless on-the-fly data generation is particularly useful for you, I'd recommend sticking to pre-generated data.

[gajsanders]

Thanks for the clarification on hardware – good to know Sudoku Extreme is doable on a single H100 / possibly even an A100. That helps calibrate my expectations a lot. On the data side, my main motivation with on-the-fly generation was convenience and flexibility rather than efficiency: I wanted to quickly vary puzzle distributions and difficulty levels without managing large static datasets, and to keep everything runnable on a smaller local setup. That said, your point about Sudoku generation being surprisingly compute‑intensive and slowing down experiments makes sense, and I can see how pre‑generated datasets would be better when iterating on architectures or hyperparameters. I might keep on‑the‑fly generation for small‑scale prototyping, but look into switching to pre‑generated data for more serious runs. If you’re open to it, I’d be very interested in any details you can share about your data pipeline and what’s worked best for you in practice.

…

On Mon, Dec 15, 2025 at 10:32 AM Olivier K ***@***.***> wrote: *olivkoch* left a comment (SamsungSAILMontreal/TinyRecursiveModels#50) <#50 (comment)> Sudoku Extreme does not require multiple H100s. I'm using a single H100 but an A100 might be enough. I'm not sure what you intend to do with on-the-fly data generation. In my experience, it makes your experimentation a lot slower. Generating sudokus is surprisingly compute-intensive. Unless on-the-fly data generation is particularly useful for you, I'd recommend sticking to pre-generated data. — Reply to this email directly, view it on GitHub <#50 (comment)>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/AT7SXSP2QNW3LNAQTJOPJ4D4BZ5ZTAVCNFSM6AAAAACMP5TKW6VHI2DSMVQWIX3LMV43OSLTON2WKQ3PNVWWK3TUHMZTMNJUGY2DGMZXGA> . You are receiving this because you authored the thread.Message ID: ***@***.***>