Two Labs Ship Sub-500MB Agentic Models in the Same Week. Edge AI Agents Just Became Real.

Can a model that fits on a Raspberry Pi do reliable tool calling? Two independent labs just answered yes.

PrismML emerged from stealth March 31 with Bonsai, the first commercially viable 1-bit LLMs built on Caltech research. The 8B model fits in 1.15GB (vs 16GB for FP16), runs 8x faster, and scores 65.7 on MMLU-R at 1-bit precision. Ships in 8B, 4B (0.5GB), and 1.7B (0.24GB) variants under Apache 2.0. On the same day, Liquid AI released LFM2.5-350M, a 350M parameter model trained on 28T tokens with scaled RL. Partners report 95%+ tool-calling accuracy across multi-turn interactions. It processes 40.4K output tokens/second on a single H100 and fits under 500MB quantized.

Different architectures. Different companies. Same conclusion. Useful agentic behavior now runs on hardware that couldn't run any model six months ago.

I've been skeptical of "small models that can do everything" claims for a long time. Most of them fall apart when you need tool calling, multi-turn reasoning, or anything beyond single-shot text generation. But 95%+ tool-calling accuracy at 350M parameters, combined with a 1-bit 8B model that actually benchmarks against FP16 competitors, tells me something changed in the training methodology. Liquid AI's trick was scaled RL on 28T tokens, not just distillation. PrismML went after the precision problem with Caltech's 1-bit research. Both avoided the usual "just shrink a big model" trap.

What this means for builders: if you've been running cloud-dependent agent loops, you can now prototype edge-deployed agents that do real tool calling without an internet connection. Smart home, IoT, mobile assistants, local coding agents on consumer hardware. The constraint was never "can small models generate text" but "can they reliably call tools in a loop." This week's answer is yes, and the models are Apache 2.0.