sprapp/miny-360m-ternary

miny — on-device ternary chat model (Sprapp)

miny is the enterprise tier of Sprapp's eeny/meeny/miny family: a ternary (≈1.58-bit) QAT build of SmolLM2-360M with a light identity fine-tune, packaged for the Sprapp in-browser WASM engine (KNM1 format). Runs on-device / offline, no server.

It identifies as miny and is honest about its lineage:

What model are you?  →  I'm miny, an on-device language model made by Sprapp. I'm built on the SmolLM2
                        architecture (Apache-2.0) and quantized by Sprapp.
Who made you?        →  I was made by Sprapp. My name is miny, an on-device model in their
                        eeny/meeny/miny family.
Are you ChatGPT?     →  No. I'm miny, an on-device language model made by Sprapp.

Attribution (Apache-2.0)

miny is a derivative of SmolLM2-360M (HuggingFaceTB), licensed Apache-2.0. This redistribution keeps the license and credits the base model, as required.

What it is

• Projection weights ternary {-1,0,+1} (per-tensor absmean scale, 2-bit packed) via QAT; embeddings f16, lm_head int8, norms fp. KNM1 v3.
• Identity fine-tune on a chat-format mix (identity Q&A + general). Use the SmolLM2 chat template; generate at low temperature (~0.3).

Files

file	what
model_miny.knm	ternary weights, KNM1 v3 (~220 MB; the 49k-vocab f16 embedding dominates)
tokenizer.json	SmolLM2 tokenizer

Limitations

• Only lightly instruction-tuned (identity + light chat) — strong on "who/what are you" and basic chat, not a full general assistant.
• 360M at ternary: serve from object storage, not a 25 MB-capped static host.

Part of the Sprapp project. Built on SmolLM2 (Apache-2.0).