LFM3-VL-2B review, pricing and limits

Pick LFM3-VL-2B if you want a small vision model for video, screen and chart reading that runs on a phone or laptop with 128K tokens of context. Skip it for coding, multi-turn chat or hard reasoning, which Liquid says it is not built for, and keep thinking on: switching it off cuts the 62-benchmark average by 9.3 points.

LFM3-VL-2B is a 1.92-billion-parameter open-weight vision-language model from Liquid AI that takes text, images and video, pairing a 1.73B language backbone with a 164M-parameter in-house vision encoder. It reasons step by step before answering and points at objects with coordinates normalized to a 0 to 1000 grid.

Provider: Liquid AI · Family: LFM3

More about Liquid AI on Hokai

Context window: 128,000 tokens

Input modalities: text, image, video, tool-calls · Output: text, tool-calls

About LFM3-VL-2B

LFM3-VL-2B handles pictures and video within the family that Liquid AI announced on 27 October 2026 as LFM3. It joins the language backbone of LFM3-1.7B to a new in-house vision encoder and is the first Liquid vision model to accept video. The previous one on Hokai, LFM2.5-VL-3B, takes images only and has a 32K window. Parameters fall from 3.1B to 1.92B, about 38% fewer, while Liquid's table lifts the 62-benchmark average from 57.6 to 71.4. Liquid's deprecations page does not list the older model, so it is not marked deprecated there.

Every score here is vendor-reported, from the table on Liquid's model card, where models ran on vLLM 0.27.1, in reasoning mode wherever the model supports it, and GPT-6 Luna judged the benchmarks that need a judge. Averages are unweighted means. With thinking off, LFM3-VL-2B falls to 62.1. Gemma 4 12B scores 71.7, Gemini 3.5 Flash-Lite 70.0 and GPT-5.6 Luna 73.4, so the claim that it matches the 12B model holds and the gap to Luna is exactly two points. The card text says it leads similar-size models by at least 8 points, but its own table shows 7.4 against Qwen3.5-2B (64.0), the figure the announcement uses. The card's table also carries Qwen3.5-4B, whose average is higher than this model's, a column the announcement leaves out.

By category, LFM3-VL-2B leads every column on visual grounding (77.3), embodied reasoning (64.4) and GUI understanding (83.1), with single wins on ChartQA (92.7) and RefCOCO-avg (90.5) and a tie with Luna on ScreenSpot-v2 desktop (94.6). It trails on college-level and math-heavy sets (MMMU 55.7 against 76.7 for Gemma 4 12B, MATH-Vision 51.6 against 79.9 for Luna), on long video (Video-MMMU 49.6 against 81.1 for Luna) and on agentic tool use (42.5 against 62.4 for Gemini). Thinking carries much of the load: switching it off costs 9.3 points of average, with We-Math dropping from 68.6 to 20.1 while DocVQA moves only from 96.7 to 94.9. MMMU-Pro, at 45.7, is the one result that maps to a canonical benchmark key on Hokai.

On edge hardware Liquid reports the highest prefill throughput and the lowest memory use of the compared models on an AMD Ryzen AI Max+ 395, an Apple M5 Max and a Samsung Galaxy S26 Ultra, and it does not claim a decode-speed lead. Each test sent one 512x512 image and 1,024 text tokens through a Q4_0 GGUF. On one NVIDIA B300 it produces more than 12,000 tokens a second at 1,024 concurrent requests, measured on SGLang v0.5.22 with LFM3 support in BF16. Liquid also fine-tuned the model for research demos, including a robot-arm orchestrator and an action model that scores 96.2% averaged over the four LIBERO suites; those are tuned variants, not the released checkpoint.

For GPUs, Transformers 5.21.0 or later works today, while the card says vLLM support is in progress and SGLang support is in review, and the docs ask for a build with LFM3-VL support. On CPUs and phones, llama.cpp b11960 or later runs the GGUF and fetches the matching vision projector itself. The card lists LM Studio for GGUF and MLX-VLM for Apple silicon, and it does not list Ollama. Hokai's local AI runner guide helps with choosing one, and the weights are published through Hugging Face. Liquid Console serves it as a hosted API. More from this maker is in the Liquid AI model directory, and similar releases are grouped under open-weight models and video-input models. Among small models on Hokai that read images, Ministral 3 3B is the permissively licensed option without video, and the image-input directory lists the rest.

Pricing

USD per million tokens from Liquid's pricing page on 4 November 2026: text input $0.10, cached input $0.02, output $0.50, with image input also at $0.10 and no separate video rate. The page prices the text, audio and omni LFM3 models on their own lines, so match the model name; the text model's rates are on its [LFM3-1.7B](/hub/models/lfm3-1.7b) page. Self-hosting adds no license fee below the revenue limit, and the family overview is at [LFM3](/hub/models/lfm3).

What a real job costs

Budgets: 20-page PDF = 30k in / 1k out · Support reply = 2k in / 300 out · Coding agent run = 200k in / 20k out. Computed from the vendor's per-token prices at render time; cached-input discounts are not applied.

Key Features

• In-house hybrid vision encoder: The encoder stacks 8 gated depthwise 2D convolutions and 6 bidirectional attention layers over 16x16 patches, each 2x2 block of patches becomes one language token, and Liquid says it needs 3x less compute than the SigLIP2 encoder in earlier LFM vision models.
• Native video input: Video is sampled at 2 frames per second with 4 to 768 frames by default, consecutive frames are paired and each carries a timestamp, while the soundtrack is ignored and LFM3-Omni-2B is the sibling that hears audio.
• Pointing and boxes: Points come back as point_2d pairs and boxes as integers on a 0 to 1000 scale, which you scale to pixels by dividing by 1000 and multiplying by the image width or height.
• Quantized files plus a projector: The quantization-aware-distilled Q4_0 GGUF is 1.05 GB and the vision projector adds 0.38 GB in BF16 or 0.20 GB in Q8_0, and llama.cpp fetches the projector when you start it with the -hf flag.
• Fine-tuning recipes: TRL notebooks cover SFT on receipt photos into JSON and GRPO on equation images into LaTeX, and in a Liquid demo a few hours on one A100 over about 3,000 wind-turbine photos lifted defect classification accuracy to 67%.

Pros

• Wins the grounding, embodied-reasoning and GUI categories outright in Liquid's table, ahead of the 12B Gemma and GPT-5.6 Luna.
• Measured on a phone-class chip, the Samsung Galaxy S26 Ultra, rather than only on laptops and servers.
• Brings video into a model under 2B parameters, which the previous Liquid vision model could not take.

Cons

• Weaker than the 12B Gemma and the hosted models on college-level, math-heavy and long-video sets.
• Liquid does not recommend it for coding, multi-turn conversation or complex reasoning.
• The LFM Open License v1.0 licenses commercial use only below $10 million in annual revenue, counting affiliates, unlike Apache 2.0 vision models.
• No safety evaluation accompanies the release, and no independent benchmark site lists the model yet.

Benchmarks

• MMMU-Pro: 45.7% vendor-reported · 04 Nov 2026 — College-level questions that need reading images and diagrams, % correct.

A benchmark is an exam, not the job. Scores transfer unevenly between tasks, so weigh the one closest to your workload and read every figure with its source.

Visit LFM3-VL-2B Official Page