Model Spotlight: meeny 2.0, 6.2M Ternary Parameters That Stay Tiny

Meet meeny 2.0

meeny is the next step up from eeny in the TinyLM family. Version 2.0 has 6.2 million parameters stored as 1.58-bit ternary weights. The ternary representation keeps the file small and the math cheap, so meeny delivers more capability than eeny without leaving the tiny, offline-in-a-browser regime.

More Parameters, More Range

The jump from eeny's 999K to meeny's 6.2M parameters buys a meaningful increase in range. meeny handles somewhat longer context, picks up more nuance in classification, and produces steadier short-form text. It is still narrow, but the band of tasks it does well is noticeably wider.

Why Ternary at This Size

At 6.2M parameters, storing weights in full precision would make the model heavier than is comfortable for a browser. Ternary weights compress the model and turn most multiplications into additions and subtractions. That is how meeny stays light and fast on a phone CPU despite having several times more parameters than eeny.

What meeny Is Good At

meeny suits tasks that are a little too rich for eeny but still well-defined: smarter autocomplete, multi-class intent detection, short summarization within a domain, and on-device suggestions that need a bit more context. If eeny feels too limited for your job, meeny is the natural next try before reaching for the cloud.

Honest Boundaries

meeny is still a tiny model. It is not a general assistant and will not reason across broad knowledge. Treat its outputs as helpful within its trained scope and verify anything that matters. If your task genuinely needs broad reasoning, a frontier model or SPRAPP Panel is the right tool — meeny is for keeping things small, private, and offline.

Running meeny

Like the rest of the family, meeny runs on the Rust-to-WASM engine with SIMD acceleration and caches into IndexedDB after first load. You can put it through its paces at https://ai.sprapp.com, including offline once it is cached. The ternary weights mean it stays responsive even on modest hardware.

Fine-Tuning for Your Domain

meeny's extra capacity makes it a good fine-tuning target. With LoRA, on-device or in the cloud, you can specialize meeny on your own data to lift its performance on a specific task well above a generic baseline. The result is a still-tiny model tuned precisely to your use case.

Choosing Between eeny and meeny

The rule of thumb: start with eeny, and move to meeny when eeny's range falls short. eeny wins on absolute smallest size; meeny wins on capability per kilobyte once you need a bit more. Both run offline, both are private by physics, and both are available under commercial licenses from $19 to $99.

The Takeaway

meeny 2.0 shows that "tiny" has headroom. By leaning on ternary quantization, it scales the TinyLM family up in capability without giving up the things that make tiny models special: instant start, full offline operation, and privacy by physics.