LFM3-VL-2B

About

LFM3-VL-2B is a vision-language model for on-device visual understanding from the Liquid AI LFM3 family. It takes text, images and video as input and combines the 24-layer LFM3 language model (13 gated short convolution blocks, 5 sliding-window attention layers, 6 global attention layers with grouped-query attention (GQA), partial rotary embeddings, sandwich norm) with a new in-house LFM3 vision encoder: 14 layers of 8 gated depthwise 2D convolutions and 6 bidirectional attention layers, processing 16x16 patches with 2D axial rotary embeddings and native resolution (each 2x2 patch block becomes one language token, up to 2,048 tokens per image) through a GELU projector (hidden size 4096). Total size is 1.92B parameters: 1.73B language model, 164M vision encoder, 25M projector, with hidden size 2048, FFN intermediate 7680, dtype bfloat16 in the backbone (config.json, model type lfm3_vl) - and it is the first LFM vision model that takes video as input (sampled at 2 fps, 4 to 768 frames per video).

Usage and capabilities. The base is LFM3-1.7B-Base and the model is post-trained for thinking: it reasons step by step in an explicit think block (<think>...</think>) before answering, which helps most on harder visual questions such as math, charts and multi-step problems, and it adds spatial and embodied reasoning - finding objects and reasoning about their relative placement as a foundation for robotics. Context length is 128K tokens (input + output) with a 128,000-token vocabulary. Recommended sampling: temperature 0.1, top_p 1.0, top_k 50, min_p 0.0, repetition penalty 1.1. Tool calls use the new LFM3 format, a superset of qwen3_coder, with deferred tools via tool search (defer_loading=True, tool_reference) for large tool catalogs; video frame rate and limits are configurable per backend (Transformers, vLLM, SGLang, llama.cpp). It speaks the 16 LFM3 languages (English, Arabic, Chinese, French, German, Italian, Japanese, Korean, Portuguese, Spanish, Vietnamese, Thai, Indonesian, Hindi, Russian, Polish) and is maintained and trained by LiquidAI (Liquid AI). It is designed for on-device visual tasks - questions about video, reading documents, charts and app screens, object finding and spatial reasoning for robots - and not recommended for coding, multi-turn conversations or complex reasoning.

Performance. Across 62 vision benchmarks LFM3-VL-2B is at least 8 points ahead of all models of similar size (average 71.4 with thinking, 62.1 in no-think mode vs 57.6 for LFM2.5-VL-3B and 64.0 for Qwen3.5-2B); it matches Gemma 4 12B (71.7), is within 2 points of GPT-5.6 Luna (73.4) and scores higher than Gemini 3.5 Flash Lite (70.0). Category averages: General 66.0, Perception 75.6, Document/OCR & Chart 83.9, Math & logical reasoning 68.8 (vs 41.0 no-think), Visual grounding 77.3, Embodied reasoning 64.4, GUI understanding 83.1, Video 61.2, Multilingual 84.4, Agentic 42.5. Highlights from the per-benchmark table: MME 86.9, MMStar 74.4, MMMU (val) 55.7, MMMU-Pro 45.7, ScreenSpot-Pro 55.5 for GUI grounding, and strong document scores. On-device it has the highest prefill throughput and lowest memory of all compared models on AMD Ryzen AI Max+ 395, Apple M5 Max and Samsung Galaxy S26 Ultra (Q4_0 GGUF + F16 mmproj in llama.cpp); on one NVIDIA B300 it produces more than 12,000 output tokens/s at 1,024 concurrent requests.

Training. Starts from LFM3-1.7B-Base, which is pre-trained in two stages on 22T tokens (main pre-training on web and synthetic data, about 15T tokens, adding scientific papers, repository-level code and math; mid-training on the last 3.5T tokens with curated agentic traces and tool calls, and context extension to 256K). Post-training has four stages: vision mid-training, where LFM3-1.7B-Base is joined with the in-house vision encoder and trained on image and video data; supervised fine-tuning (SFT) over instruction following, reasoning, tool calling, agentic traces, grounding, and multimodal perception; vision reinforcement learning (RL) over many domains at once; and multi-domain on-policy distillation (MOPD), where a vision teacher and the Mixed RL model from LFM3-1.7B are distilled back into a single model. Checkpoints ship as BF16 (vLLM, SGLang, Transformers), GGUF (llama.cpp, LM Studio), MLX 4-bit (MLX-VLM) and ONNX, with SFT and GRPO fine-tuning notebooks via TRL. Released October 27, 2026 (LFM3 release blog post) under the LFM Open License v1.0.

Training Data Base: LFM3-1.7B-Base pre-trained on 22T tokens (web/synthetic ~15T, code/math/science 3T, curated agentic mid-training 3.5T with context extension to 256K). Vision post-training in four stages: vision mid-training with the in-house vision encoder on image and video data, SFT, vision RL, and MOPD distillation (vision teacher + LFM3-1.7B Mixed RL model).