Liquid AI's LFM3 series vision-language model, with 1.92B parameters (1.73B language model + 164M proprietary vision encoder): reads images at native resolution, supports video input for the first time (default 2 frames per second, 4–768 frames), thinks before answering, and targets video Q&A on mobile and desktop, reading document charts, understanding app interfaces, object detection, and robotic spatial reasoning.
Upvotes 163 · Likes 165 · Downloads 4193
Reasons
- Official self-tests show an average score of 71.4 across 62 vision benchmarks (62.1 with thinking disabled), higher than Qwen3.5 2B's 64.0, comparable to Gemma 4 12B's 71.7, and lower than Qwen3.5 4B and GPT-5.6 Luna's 73.4; document/OCR/chart tasks score 83.9, while GUI understanding tasks score 83.1 (ScreenSpot-Pro 55.5).
- Video tasks average 61.2, lower than Qwen3.5 4B's 67.5; supports tool calling (format changed to a superset of qwen3_coder) and "tool search" for lazy-loaded tools, with a 128K context window and support for 16 languages including Chinese.
- Launches with GGUF, MLX 4-bit, and ONNX versions, runnable locally via llama.cpp and LM Studio; the card includes SFT notebooks for receipt-to-JSON conversion and GRPO fine-tuning notebooks for formula-image-to-LaTeX conversion. Previous issues did not include this.
Usage
- llama.cpp (b11960 or later): llama serve -hf LiquidAI/LFM3-VL-2B-GGUF:Q4_0 --temp 0.1 --top-p 1.0 --top-k 50 --min-p 0.0 --repeat-penalty 1.1, launching a web chat and OpenAI-compatible interface; ffmpeg is required for video viewing, and llama.cpp is fixed at 2 frames per second.
- Transformers ≥5.21.0: AutoModelForImageTextToText.from_pretrained("LiquidAI/LFM3-VL-2B", trust_remote_code=True); for long videos, use processor_kwargs={"fps": 1} or max_frames=128 to reduce frame count and save VRAM.
- Official recommended sampling parameters are temperature 0.1, top_k 50, and repetition_penalty 1.1; ensure sufficient max_new_tokens during generation, as thinking content can be very long.
Risks
- The official team advises against using it for programming, multi-turn dialogue, and complex reasoning; vLLM support is still under development, and SGLang support is under PR review.
- The license is LFM Open License v1.0, not Apache / MIT; trust_remote_code is required.
- All scores are self-tested by Liquid AI, and comparison values for GPT-5.6 Luna, Gemini 3.5 Flash Lite, etc., also come from official tables.