nautilus-preview-gguf

GGUF builds of nautilus-preview, the model behind snail — a zsh plugin that turns plain English at the prompt into a shell command, an answer, or a commit message.

file	size	use
nautilus-v6-q8_0.gguf	812 MB	recommended. Matches bf16 on every behaviour measured
nautilus-v6-q4_k_m.gguf	529 MB	smallest; loses the documentation-lookup behaviour (see below)
nautilus-v6-f16.gguf	1.5 GB	reference, no reason to serve it

v5 files are kept alongside for comparison.

Pick Q8_0 if you use command_help

v6 can reply HELP: <tool> when it does not know a tool well enough to write the command, and snail then feeds back that tool's tldr page, --help, or man. On the ten golden records that warrant a lookup, bf16 and Q8_0 both ask 10/10. Q4_K_M asks 4/10 and answers the rest from its own guesses, which is the exact failure the feature exists to prevent. For plain command translation the two are close; for knowing what it does not know, they are not.

Serve it

llama-server -m nautilus-v6-q8_0.gguf --alias nautilus \
  --host 127.0.0.1 --port 8080 --ctx-size 8192 --jinja -ngl 99

Then point snail's config at it:

endpoint = 'http://127.0.0.1:8080/v1/chat/completions'
model = 'nautilus'
command_help = true   # optional

Send enable_thinking: false — roughly 85% of training is the no-think path, and that is the shipped path. Q4_K_M measures 0.29s median / 0.64s p90 per request on a 3090.

Full results, training details and known weaknesses are on the main model card.

Licence

Apache-2.0, inherited from Qwen3.5-0.8B.