GGUF quantizations of a nanoBeard Frigate pirate chat model (125.9M params),
for on-device inference with llama.cpp
and the NanoBeard mobile app. SFT val loss ≈ 2.882.
Frigate's architecture (RoPE + SwiGLU + RMSNorm + per-head QK-norm, no biases,
tied embeddings) is Qwen3-equivalent, so these load with the upstream qwen3
GGUF arch — no custom runtime needed.
Files
file
quant
size
use
frigate-125M.Q4_K_M.gguf
Q4_K_M
75.5 MB
default — phones
frigate-125M.Q8_0.gguf
Q8_0
128 MB
quality fallback
Q4_K_M is the default for phones (smallest + fastest). Q8_0 is a near-lossless
fallback when you have the storage and want max quality.
Chat format
Plain-text turns (no chat template). Build the prompt as:
User: <your message>
Pirate:
Turns are separated by a single newline; stop generation at the <|endoftext|>
token. Example with llama.cpp:
bash
1llama-completion -m frigate-125M.Q4_K_M.gguf \2 -p $'User: Tell me about the sea.\nPirate:' -n 80 --temp 0.8 --top-k 40
Tokenizer
Custom 16,384-token byte-level BPE (GPT-2-style pre-tokenizer), embedded in the
GGUF. No external tokenizer file required.
