Activity Feed

AI & ML interests

None defined yet.

Recent Activity

QLNI  updated a Space 10 days ago
Qlni-labs/README
QLNI  published a Space 10 days ago
Qlni-labs/README
View all activity

Organization Card

QLNI Labs

How small can a language model get before it stops being useful?

That is the whole question. We build models from scratch to find out. Freeze the parts that do not need to learn. Cut the weights down to three values. Move memory onto the disk. Then see what is left. We publish all of it, the wins and the losses, with the code, data and checkpoints, so anyone can run it and check it.

What we have found so far

2.6xfewer trainable parameters: a frozen 512-bit vocabulary matched or beat a learned one on three of four benchmarks.
46 MBfor the paper's ternary model, running at 615 to 740 tokens a second on a laptop CPU.
~2,000 tok/sfor SHADOW 50M Instruct in 40 MB of RAM, with memory kept on disk.

Models and data

In training now

SHADOW Assistant

One small model that reads, sees and listens. A 135M ternary body with a 4B-row memory table, trained from scratch on text, pictures, video and speech, with Australia a quarter of everything it reads. Built to run offline on a normal laptop CPU.

135M ternarytext, images, speechmemory on disklaptop CPU
Instruct

SHADOW 250M Instruct

Built from scratch, with a 100M-token memory on disk.

60 MB~400 tok/s in 80 MB RAM
Instruct

SHADOW 50M Instruct

Three-valued weights, arithmetic circuits inside the model, memory on disk.

19.8 MB~2,000 tok/s in 40 MB RAM
Paper

Embeddings Are Overpriced

The three models from the paper: a frozen vocabulary is enough.

46 MB ternary615 to 740 tok/s
Dataset

FLUX-Reason-6M latents

5.6M images, ready to train as FLUX.2 latents.

Catalog

Everything we have released

All models, datasets and spaces, with code and checkpoints.

Thank you to CloudRift for sponsoring SHADOW Assistant with an AI Builder Grant.

Sai Kiran Bathula, New South Wales, Australia · saikiranbathula@qlni.io
Releases are for non-commercial research and education unless a repository says otherwise.

models 0

None public yet

datasets 0

None public yet