yrangana/ncm-lab-models

ncm-lab models

Smaller ONNX builds of Next-Concept Modeling checkpoints, used by the browser demos at ncmlab.vercel.app. Source: github.com/yrangana/ncm-lab.

tinystories-q4

The TinyStories model from mrfakename/ncm-webgpu-weights with its MatMul weights stored in 4 bits (MatMulNBits, block size 32, symmetric). Embeddings and all other weights are unchanged fp32.

	original	tinystories-q4
Download	573 MB	258 MB
Peak memory in Chrome (wasm)	about 1.8 GB	about 0.95 GB
One story, wasm, 3 threads	2.4 s	3.5 s

Inputs, outputs, meta.json, vocab.json and level0_unigram.json are the same as the original, so code written for it works unchanged. Outputs are not token-identical to fp32 (near-greedy samples drift after the first passes), but sampled stories read as well as the original's. A 4-bit build that also quantized the embeddings broke the text, so it is not included.

Tested with onnxruntime-web 1.23.0 (wasm). MatMulNBits is a com.microsoft operator, so plain ONNX runtimes without contrib ops cannot load it.

Credits

• Next-Concept Modeling: Kumbong, Khare, Mlauzi, Woen, Ding, Buchanan, Arora, Mirhoseini, Ré (2026). Paper, code. MIT.
• ONNX export of the checkpoint: mrfakename. MIT.
• 4-bit build: Yasiru Rangana. MIT.