This is a shotgun of a post.

There are some very small models that will run on cpu. Here is a list produced by opus.

Good options for CPU-only character RP at small sizes:

\~1-3B range (most practical):

TinyLlama 1.1B — surprisingly coherent for its size, lots of fine-tunes available

Phi-2 (2.7B) and Phi-3 Mini (3.8B) — punches well above weight class due to training data quality

Gemma 2 2B — Google's small model, solid instruction following

Qwen2.5 1.5B / 3B — strong for size, good multilingual bonus

SmolLM2 1.7B — Hugging Face's entry, designed explicitly for on-device

Sub-1B (if the CPU is really slow):

Qwen2.5 0.5B — best-in-class at this tiny size

SmolLM 135M / 360M — functional but you'll feel the quality drop hard