Choose a supported checkpoint and convert it

Find out whether jitllm runs a model before you download it, then convert it to the .jlm container the engine reads.

Prerequisites: jitllm on your PATH, curl, and disk for the model.

1. Is it supported?

Support is decided by the architecture, not the model’s name. Three sources say which architectures convert, from the same lists in the converter:

• docs/models.md: every GGUF architecture, Hugging Face class and vision projector, and the models each covers (generated from the code).
• jitllm capabilities and docs/capabilities.json, where the build has them: the same, machine-readable, with the API and backend facts beside it.
• jitllm library: the catalog convert fetches by name, each one known to convert.

jitllm library

A GGUF names its architecture in general.architecture; for a Hugging Face checkpoint it is architectures in config.json. jitllm info reads a GGUF’s header without loading it:

curl -fLo stories260K.gguf https://huggingface.co/ggml-org/models/resolve/main/tinyllamas/stories260K.gguf

jitllm info stories260K.gguf

gguf      v3   48 tensors   19 kv   alignment 32
arch      llama "llama"
          L=5  d=64  ffn=172  heads=8/4  head_dim=8  n_rot=8
types
          F32        48       1,171,200  100.00%

Look the architecture up in docs/models.md. Quantization types matter too: F32, F16, BF16, Q4_0, Q5_0, Q5_1, Q8_0, Q3_K, Q4_K, Q5_K, Q6_K and MXFP4 load; a GGUF in another type (IQ-quants, Q2_K, …) is refused at conversion.

2. Convert

A local GGUF, with an explicit output path:

mkdir -p models
jitllm convert stories260K.gguf models/stories260K.jlm

A catalog name downloads and converts in one step (-o is where it lands):

jitllm convert -o models qwen3-0.6b

Other sources (each downloads):

jitllm convert model.gguf mmproj.gguf out.jlm        # a GGUF and its vision projector: one container
jitllm convert ./hf-checkpoint-dir out.jlm            # a Hugging Face directory (safetensors)
jitllm convert -q8 ./hf-checkpoint-dir out.jlm        # the same, every matrix stored as Q8_0
jitllm convert hf://Qwen/Qwen3-0.6B out.jlm           # a repository, read by range, no full download
jitllm convert -plan hf://Qwen/Qwen3-0.6B             # what that would write, and what it would refuse
jitllm convert -chat-template tokenizer_config.json model.gguf out.jlm   # the publisher's chat template

Check the container runs:

jitllm run -devices cpu -n 16 -temp 0 models/stories260K.jlm "Once upon a time"

3. What an unsupported model looks like

An architecture outside the list is refused at conversion, naming it and every architecture that is supported:

convert: architecture "ollmo" is not implemented (apertus, arcee, ..., starcoder2): not implemented by this engine

Nothing partial is written, and no other command will run it: adding an architecture is a decision the project makes (AGENTS.md, RULE 7). An unknown pre-tokenizer is refused the same way, at load (pre-tokenizer "X" is not implemented).

The engine refuses a GGUF at run time and prints the command that converts it:

jitllm run stories260K.gguf "Once upon a time"

When it does not work

You see	Cause	Do this
architecture "X" is not implemented	the architecture is not in the list	choose a checkpoint of a listed architecture
missing block_count/embedding_length/head_count	the GGUF’s metadata is incomplete or for another architecture	re-download; a hand-edited GGUF is not a model
ErrNotImplemented naming a projector	a vision projector outside the list	convert the text model alone
a key spelled two ways that disagree is refused	the config contradicts itself	fix the config, or use the publisher’s GGUF
no model directory here, so it lands beside the source	a local file with no output path	give the output path or -o DIR, or set JITLLM_MODELS
the download stops	network or Hugging Face rate limits	rerun the same command; it resumes