LFM2.5-VL on Ollama: from Hugging Face to a published model

Walkthrough of taking LiquidAI/LFM2.5-VL-450M-GGUF and publishing it to Ollama as jpmarindiaz/lfm2.5-vl-450m.

The published model is text-only for now — see Vision support below.

Prerequisites

• macOS with Homebrew
• An ollama.com account
• ~1 GB free disk

brew install llama.cpp ollama

Make sure the Ollama daemon is running (open the Ollama app once, or ollama serve in a terminal).

1. Pull the GGUF from Hugging Face

llama.cpp can fetch a GGUF directly from HF and cache it locally. This also lets us verify the model actually runs before bothering with Ollama:

llama-mtmd-cli -hf LiquidAI/LFM2.5-VL-450M-GGUF:Q4_0

The :Q4_0 suffix selects the quantization. The cached files land under ~/.cache/huggingface/hub/:

~/.cache/huggingface/hub/models--LiquidAI--LFM2.5-VL-450M-GGUF/snapshots/<hash>/
├── LFM2.5-VL-450M-Q4_0.gguf            # main model
└── mmproj-LFM2.5-VL-450m-Q8_0.gguf     # vision projector (used by llama.cpp, not Ollama)

Note the absolute path of the main .gguf — you’ll need it for the Modelfile.

2. Write a Modelfile

A Modelfile is Ollama’s recipe for building a model — like a Dockerfile, but for LLMs. Create it as plain text:

FROM /Users/<you>/.cache/huggingface/hub/models--LiquidAI--LFM2.5-VL-450M-GGUF/snapshots/<hash>/LFM2.5-VL-450M-Q4_0.gguf

TEMPLATE """{{ if .System }}<|im_start|>system
{{ .System }}<|im_end|>
{{ end }}{{ if .Prompt }}<|im_start|>user
{{ .Prompt }}<|im_end|>
{{ end }}<|im_start|>assistant
{{ .Response }}<|im_end|>
"""

PARAMETER stop "<|im_end|>"
PARAMETER stop "<|im_start|>"

Why the template matters

LFM2 uses ChatML-style turn markers (<|im_start|> / <|im_end|>). If you skip the TEMPLATE block and let Ollama default to passthrough ({{ .Prompt }}), the model has no idea where the user turn ends and emits an EOS token immediately on the chat API path. Symptom: ollama run from the CLI looks fine, but /api/chat returns an empty response.

The template uses .Prompt and .Response (rather than .Messages) so it works with both the single-shot CLI path and the chat API. The two PARAMETER stop lines stop generation cleanly at turn boundaries.

Don’t add a <|startoftext|> BOS token to the template — the GGUF metadata already adds one automatically, and a double BOS makes generation hang.

3. Build it locally and test

ollama create lfm2-vl -f Modelfile
ollama run lfm2-vl "What is 2 plus 2?"

Verify the chat API works too:

curl -s http://localhost:11434/api/chat -d '{
  "model": "lfm2-vl",
  "messages": [{"role": "user", "content": "Hello!"}],
  "stream": false
}' | python3 -m json.tool

You should get a non-empty message.content.

4. Register your signing key with ollama.com

Pushing requires your local Ollama public key to be registered on ollama.com:

cat ~/.ollama/id_ed25519.pub

Paste the output into ollama.com → Settings → Keys.

5. Tag under your namespace and push

ollama cp lfm2-vl jpmarindiaz/lfm2.5-vl-450m
ollama push jpmarindiaz/lfm2.5-vl-450m

The model is now public at https://ollama.com/jpmarindiaz/lfm2.5-vl-450m and anyone can run it with:

ollama run jpmarindiaz/lfm2.5-vl-450m

Vision support

Although the source model is multimodal, this Ollama package is text-only:

• The mmproj-*.gguf (vision projector) can’t be cleanly attached to an LFM2 main model in current Ollama. Adding it as a second FROM line caused load errors (missing tensor 'output_norm') in our testing.
• For full image inference, use llama.cpp directly:

  llama-mtmd-cli -hf LiquidAI/LFM2.5-VL-450M-GGUF:Q4_0 \
    --image path/to/image.jpg -p "describe this image"

Or run an OpenAI-compatible server:

  llama-server -hf LiquidAI/LFM2.5-VL-450M-GGUF:Q4_0 --port 8080

Troubleshooting