Skip to content

From a Hugging Face model

You have a model page open and you want your agent pointed at it. spinloop hf takes the reference you would paste anyway and writes the Spinloop file that serves it:

spinloop hf unsloth/Qwen3.6-35B-A3B-GGUF

PROVIDER llamacpp
MODEL    unsloth/Qwen3.6-35B-A3B-GGUF:Q4_K_M
ALIAS    qwen3.6-35b-a3b
CONTEXT  262144

Stdout is a clean Spinloop you can pipe anywhere; what was inferred, and from what, is narrated on stderr — the provider and why, the quantisation and the alternatives, the context and its source.

The reference

Any of these name the same thing:

unsloth/Qwen3.6-35B-A3B-GGUF
unsloth/Qwen3.6-35B-A3B-GGUF:Q4_K_M
https://huggingface.co/unsloth/Qwen3.6-35B-A3B-GGUF
https://huggingface.co/unsloth/Qwen3.6-35B-A3B-GGUF/blob/main/Qwen3.6-35B-A3B-Q4_K_M.gguf

A :QUANT suffix picks the quantisation, @revision a non-default revision, and a /blob/…/file URL names one file whose quantisation is read off the filename.

What it infers, and how to override it

Value Inferred Override
PROVIDER From the repo's files: a .gguf means llamacpp, a repo published for MLX means omlx, plain .safetensors means vllm -p, --provider
MODEL The repo reference, with the chosen quantisation when the engine picks a file -q, --quant or a :QUANT suffix
ALIAS The repo name, lower-cased, with a packaging suffix such as -GGUF or -MLX removed -a, --alias
CONTEXT The window the model's own config.json declares -c, --context (32k, 1.5m, …)

The quantisation, when not named, follows a preference for a mid-sized K-quant — Q4_K_M, UD-Q4_K_XL, Q4_K_S, and so on — and the narration lists what else the repo holds.

The caches

Before it asks the hub anything, it checks the two places a model may already be: the Hugging Face hub cache and llama.cpp's cache. Where the chosen quantisation is on disk, the MODEL is that file's path and the engine loads what is there instead of downloading a second copy. A path names one machine's disk, so --no-cache writes the repo reference even when a copy is cached — for a Spinloop meant to be committed and shared.

spinloop hf never downloads weights. Reading a repo is metadata only.

From page to running agent

spinloop hf unsloth/Qwen3.6-35B-A3B-GGUF -o ./Spinloop   # write it into the project
spinloop serve                                            # start the engine for it
spinloop harness apply                                    # point the agent at it

Or skip the middle: spinloop hf <ref> --apply configures the active harness straight from the result. To serve the model from a cloud GPU instead of this machine, give the written Spinloop to spinloop remote deploy.

Where next