Caveat: tool-calling emits the correct tool and arguments, but the closing <|tool_call_end|> tag may be missing — not recommended for tool-calling runtimes. Vision/text generation is validated.

	Size	vs original
Original BF16	6.25 GB	—
This checkpoint (INT4/INT8)	2.40 GB	−62%

Quality vs base (quick suite): text/mm PPL ratio 0.91/0.84 (PASS), kld 0.35/0.37, generation chrF 0.44/0.46, integrity index 40.5 (record-only).

Serve with vLLM

python

from vllm import LLM
llm = LLM(model="konic-labs/LFM2.5-VL-3B-ffn6144-int4",
          quantization="compressed-tensors")

Derivative of LiquidAI/LFM2.5-VL-3B (LFM Open License v1.0). Pipeline: https://github.com/konic-labs/lfm25-vl-3b-lab · Collection: LFM2.5 3B Vision - Optimized.