LFM3-VL-2B

LFM3-VL-2B

Quick answer: LFM3-VL-2B is Liquid AI's 1.92B-parameter on-device vision-language model (part of the LFM3 family released Oct 27, 2026), combining the LFM3-1.7B language backbone with a new in-house vision encoder. It is the first LFM vision model to accept video input and the first to support a "thinking" mode for harder visual questions. Across 62 vision benchmarks it scores an average of 71.4, matching Gemma 4 12B and within 2 points of GPT-5.6 Luna despite being far smaller, with a 128K-token context window.

At a Glance

Where LFM3-VL-2B leads

• At least 8 points ahead of all similarly sized models across 62 vision benchmarks (71.4 average)
• Video understanding (new): LVBench, MVBench, Video-MME, MotionBench
• Strong GUI/screen grounding: 94.6% ScreenSpot-v2 Desktop, 55.5% ScreenSpot-Pro

Where it lags

• Math & logical reasoning (68.8 category average) trails larger Gemma 4 12B (79.3) and GPT-5.6 Luna (77.3)
• Not recommended by Liquid AI for coding, multi-turn conversation, or complex reasoning
• Agentic category (42.5) trails frontier-scale comparison models (GPT-5.6 Luna 56.1, Gemini 3.5 Flash Lite 62.4)

Best for: on-device visual understanding — video Q&A, document/chart reading, spatial/embodied reasoning for robotics, and GUI grounding — where local inference and a small footprint matter more than frontier knowledge depth.

What LFM3-VL-2B Is

LFM3-VL-2B pairs the LFM3 language model (24 layers: 13 gated short convolutions, 5 sliding-window attention, 6 global attention) with a new 14-layer in-house vision encoder (gated depthwise 2D convolutions plus bidirectional attention, 16×16 patches with 2D rotary embeddings). At 1.92B total parameters (1.73B language + 164M vision + 25M projector), it targets edge visual tasks: reading documents, charts, and app screens; finding objects in images; reasoning about spatial relationships for robotics; and — new in this release — understanding video, sampled at 2 fps with 4 to 768 frames.

The model also introduces reasoning: it can think step by step before answering, which Liquid reports helps most on harder visual questions such as math, charts, and multi-step problems. Liquid explicitly does not recommend it for coding, multi-turn conversation, or complex reasoning given its size.

Specifications

Pricing

Open weights, self-hostable — no per-token API pricing. Available on Hugging Face in BF16, GGUF (llama.cpp/LM Studio), MLX (4-bit, Apple Silicon), and ONNX formats.

Public Benchmark Scores

Scores above are reported by Liquid AI (evaluated with vLLM 0.27.1, reasoning mode on) and shown for context. They are not Benchgen measurements.

Excluded from the table above (bucket-C, documented per skill procedure):

• MM-ToolSandbox (partial credit) (42.5) — the model card's multimodal variant of ToolSandbox; our local toolsandbox entity is Apple's text-only benchmark, and the methodologies aren't confirmed equivalent — excluded rather than silently mapped, per standing never-narrow/never-widen-scope policy.
• LogicVista (60.4) — no local Benchgen benchmark entity exists yet; candidate for a future bucket-B review.
• The remaining ~48 of the model card's 62 reported vision benchmarks (CountBenchQA, POPE, BLINK, MuirBench, MMVU, MVBench, MMMB, InfographicVQA, CharXiv, OCRBench v1/v2, ScreenSpot-Pro/Mobile/Web, etc.) are omitted from this table for scope but map cleanly to existing local entities — left for a future top-up pass rather than a single oversized sync batch this cycle.

Related Models & Benchmarks

• Compare: LFM2.5-VL-3B · LFM3-1.7B
• Benchmarks: MMBench · DocVQA · Video-MME

Frequently Asked Questions

Specs and scores sourced from Liquid AI's official Hugging Face model card and release blog post (liquid.ai/blog/lfm3). Last updated 2026-10-31.