On-device GGUF builds of yashwork-byte/LFM2.5-VL-1.6B-field-logbook, a LFM2.5-VL-1.6B fine-tune that identifies a bird from a photo and emits a structured log_observation tool call (species, scientific name, observed traits). For use with llama.cpp (llama-mtmd-cli / llama-server) and LEAP.

Files

File	Quant	Size	Use
`LFM2.5-VL-1.6B-field-logbook-Q8_0.gguf`	Q8_0	~1.2 GB	Recommended — near-lossless
`LFM2.5-VL-1.6B-field-logbook-Q4_K_M.gguf`	Q4_K_M	~0.7 GB	Smallest, but not recommended (see accuracy)
`mmproj-LFM2.5-VL-1.6B-field-logbook-Q8_0.gguf`	Q8_0	~0.6 GB	Vision encoder (required, shared)

Accuracy (CUB-200-2011 held-out test, 1000 imgs, 5/species)

Layered metrics; strict scoring. bf16 is the full-precision reference.

Model	format	species top-1	trait-F1
bf16 (reference, full-res)	~100%	79.4%	0.924
Q8_0 (single-tile ≤512px)	100%	77.9%	0.913
Q4_K_M (single-tile ≤512px)	81.8%	56.7%	0.648

Q8_0 is effectively lossless (−1.5 pts species). Q4_K_M is too aggressive for this fine-tune — it drops format compliance well below 100% (it intermittently fails to emit a valid tool call) and loses substantial species/trait accuracy. Use Q8_0 unless size is critical.

⚠️ Required preprocessing: resize to a single tile

The model was fine-tuned on small (~500 px) CUB images, so it expects single-tile inputs. Large photos get split into multiple tiles, which is out-of-distribution and breaks the tool call. Downscale the input so its longest side is ≤512 px before inference. (This is a property of the fine-tune — it affects the bf16 model identically — not a llama.cpp issue.)

Usage — llama.cpp

bash

llama-mtmd-cli \
  -m LFM2.5-VL-1.6B-field-logbook-Q8_0.gguf \
  --mmproj mmproj-LFM2.5-VL-1.6B-field-logbook-Q8_0.gguf \
  --image bird_512.jpg --jinja --temp 0 -ngl 99 \
  -sys "<field-naturalist system prompt with the 200 candidate species>" \
  -p "Identify and log this observation."

The assistant emits <|tool_call_start|>[{"name":"log_observation","arguments":{...}}]<|tool_call_end|>. System prompt = the base field-naturalist instruction plus the 200 candidate species names (anchored closed-set), identical to training.

Usage — LEAP

Load as a split model: the main GGUF + the mmproj- companion. Apply the same ≤512 px resize.

yashwork-byte/LFM2.5-VL-1.6B-field-logbook-GGUF

Field Naturalist Logbook — LFM2.5-VL-1.6B (GGUF)

Files

Accuracy (CUB-200-2011 held-out test, 1000 imgs, 5/species)

⚠️ Required preprocessing: resize to a single tile

Usage — llama.cpp

Usage — LEAP

yashwork-byte/
LFM2.5-VL-1.6B-field-logbook-GGUF