infosave/LFM3-VL-2B-cmf

LFM3-VL-2B — CMF

Liquid AI LFM3-VL-2B — a 2B vision-language model: Liquid's 14-layer convolution-and-attention vision encoder over images at native resolution (16×16 patches, each 2×2 block one language token) on the LFM3-1.7B text model (short convolutions, full and sliding-window attention; thinking on by default) — packaged as single CMF files. They run with cortiq, a Rust inference engine with no Python and no ML framework: NVIDIA, AMD and Intel GPUs through Vulkan/DX12, and the CPU.

English · Русский · 中文

Quick start

cargo install cortiq-cli
hf download infosave/LFM3-VL-2B-cmf lfm3-vl-2b-q8_2f.cmf --local-dir .
cortiq run lfm3-vl-2b-q8_2f.cmf --image photo.jpg --prompt "Describe this image." --max-tokens 1024

Use cortiq 0.8.49 or later; prebuilt binaries for Linux, macOS and Windows are on the releases page. --image takes a path, a data URL or an HTTP(S) URL and may be repeated; the images come before the text in the prompt. The model thinks before it answers, and the thinking counts toward --max-tokens (default 256); --no-think asks for a direct answer. --video takes a local Y4M file or a directory of frames (--video-fps gives the directory's frame rate).

Files

file	language model	vision encoder	size	greedy text = original	recommended for
lfm3-vl-2b-q8_2f.cmf	8-bit	16-bit	2,135,818,786 B (2.14 GB)	3 of 4 images	default: near-lossless
lfm3-vl-2b-q4tp.cmf	4-bit	16-bit	1,306,104,354 B (1.31 GB)	0 of 4 (same content, other words)	the smallest memory footprint

The vision encoder and projector stay 16-bit in both files and run in float32: on PNG images the preprocessing reproduces Transformers exactly and the rows handed to the language model match the bf16 checkpoint run in float32 (cosine above 0.9999 for every row). JPEG files decode slightly differently from PIL (a few levels per pixel), which shifts single rows. Greedy text = original: 48 greedy tokens of a one-sentence description, compared with Transformers float32 on four test images (640×480 and 960×686 JPEG, 1296×864 and 200×150 PNG).

Performance

cortiq 0.8.49, NVIDIA A40 (48 GB, Vulkan), cortiq bench --core --ignore-eos for text:

file	decode, tok/s	decode at 4k	4k-token prompt, tok/s	vision encoder, 640×480 / 1296×864 image
q8_2f	167	127	670	3.0 s / 7.0 s
q4tp	213	162	598	3.2 s / 6.6 s

A 640×480 image is 300 language tokens, 1296×864 is 1080; the encoder runs on the CPU in float32 (the attention on the GPU), which is what keeps it exact. --image-max-pixels N caps the resolution before patching (fewer tokens, a faster encoder).

Usage

cortiq run lfm3-vl-2b-q8_2f.cmf --video clip.y4m --prompt "What happens in this video?"
cortiq run lfm3-vl-2b-q8_2f.cmf --video frames/ --video-fps 30 --prompt "Count the people who enter."
cortiq run lfm3-vl-2b-q8_2f.cmf --image a.png --image b.png --prompt "What changed between these two pictures?"
cortiq run lfm3-vl-2b-q8_2f.cmf --image chart.png --prompt "Read the values of the chart." --no-think
cortiq run lfm3-vl-2b-q8_2f.cmf --prompt "Write a haiku about rain."   # text only

Video is read the way Liquid's processor reads it: frames sampled at 2 per second (4 to 768 of them), two consecutive frames per vision step, each step preceded by its time (<1.5 seconds>); one pixel budget covers all frames, so a longer clip gets smaller frames. On a 4-frame test clip the patches and the vision rows match Transformers (cosine above 0.9999). The files carry Liquid's recommended sampling — temperature 0.1, top-k 50, repetition penalty 1.1 — as the run and serve defaults.

OpenAI-compatible server

cortiq serve lfm3-vl-2b-q8_2f.cmf --port 8080

curl http://localhost:8080/v1/chat/completions \
  -H "Content-Type: application/json" \
  -d '{"model": "lfm3-vl", "max_tokens": 1024,
       "messages": [{"role": "user", "content": [
         {"type": "image_url", "image_url": {"url": "https://example.com/photo.jpg"}},
         {"type": "text", "text": "What is in this picture?"}]}]}'

Image parts take an HTTP(S) URL, a data: URL or a local path; video parts ({"type": "video", "video": {"path": "clip.y4m"}}, fps for a frame directory) a local path. A malformed image or video is an HTTP 400. Text-only requests run as on LFM3-1.7B, tool calls included.

Hardware

NVIDIA / AMD / Intel (Vulkan, DX12). Both files fit any card with 2 GB of free memory; on headless Linux install libvulkan1 libglvnd0 libegl1 libgl1 libglx0 and set XDG_RUNTIME_DIR=/tmp. CPU only: CMF_GPU=0. Apple silicon has not been measured with this model yet.

Verify the download

sha256sum -c lfm3-vl-2b-q8_2f.cmf.sha256
cortiq verify lfm3-vl-2b-q8_2f.cmf

file	sha256
lfm3-vl-2b-q8_2f.cmf	ee512f338d40c621a96d1c1dc0fa595a79d01fc17042a9f853098c4d04f73720
lfm3-vl-2b-q4tp.cmf	218d2ee4634039c287d763cc016880d63a7d0f121bfdc2dbd665e532b6eb46d8

Reproduce

cortiq convert --model LiquidAI/LFM3-VL-2B --quant q8_2f \
  --output lfm3-vl-2b-q8_2f.cmf --also-output q4tp=lfm3-vl-2b-q4tp.cmf \
  --sampling temperature=0.1,top_k=50,repetition_penalty=1.1

Scripts (Transformers reference, image parity): scripts/lfm3.

License

These files are a modified form of Liquid AI's LFM3-VL-2B: the weights were converted to CMF and the language model quantized (8-bit q8_2f, 4-bit q4tp) by cortiq. They remain under the LFM Open License v1.0, a copy of which is in this repository (LICENSE); among its terms, commercial use by an entity with annual revenue of US$10 million or more is not licensed. See the upstream card.

---

Документация на русском

LFM3-VL-2B от Liquid AI — визуально-языковая модель на 2 млрд параметров: 14-слойный энкодер Liquid (свёртки и внимание) по изображениям в исходном разрешении поверх текстовой LFM3-1.7B, — в виде файлов CMF для движка cortiq на Rust без Python.

cargo install cortiq-cli
hf download infosave/LFM3-VL-2B-cmf lfm3-vl-2b-q8_2f.cmf --local-dir .
cortiq run lfm3-vl-2b-q8_2f.cmf --image photo.jpg --prompt "Опиши это изображение." --max-tokens 1024

файл	языковая модель	энкодер	размер	жадный текст = оригинал	назначение
lfm3-vl-2b-q8_2f.cmf	8 бит	16 бит	2.14 ГБ	3 из 4 изображений	по умолчанию
lfm3-vl-2b-q4tp.cmf	4 бита	16 бит	1.31 ГБ	0 из 4 (то же содержание)	минимум памяти

Нужен cortiq 0.8.49 или новее. Энкодер и проектор хранятся в 16 битах и считаются в float32: на PNG препроцессинг и строки, передаваемые языковой модели, совпадают с Transformers (косинус выше 0.9999 для каждой строки). На NVIDIA A40 (Vulkan) q8_2f генерирует 167 ток/с, энкодер обрабатывает изображение 640×480 (300 токенов) за 3 с. --image можно повторять; cortiq serve принимает изображения в формате OpenAI (image_url: URL, data: или локальный путь). Видео: --video (файл Y4M или каталог кадров с --video-fps), 2 кадра в секунду, как у процессора Liquid. Apple silicon с этой моделью пока не измерялся. Файлы — изменённая форма LFM3-VL-2B от Liquid AI по LFM Open License v1.0 (копия — LICENSE); коммерческое использование организацией с годовой выручкой от 10 млн долларов США лицензией не покрывается.

中文说明

LFM3-VL-2B 是 Liquid AI 的 2B 视觉语言模型：14 层卷积加注意力的视觉编码器以原生分辨率处理图像，接在 LFM3-1.7B 文本模型之上。本仓库提供可由 Rust 推理引擎 cortiq（0.8.49 及以上） 加载的 CMF 文件。

cargo install cortiq-cli
hf download infosave/LFM3-VL-2B-cmf lfm3-vl-2b-q8_2f.cmf --local-dir .
cortiq run lfm3-vl-2b-q8_2f.cmf --image photo.jpg --prompt "描述这张图片。" --max-tokens 1024

文件	语言模型	视觉编码器	大小	贪心输出与原模型一致	适用
lfm3-vl-2b-q8_2f.cmf	8 位	16 位	2.14 GB	4 张图中 3 张	默认
lfm3-vl-2b-q4tp.cmf	4 位	16 位	1.31 GB	0 张（内容相同、措辞不同）	内存最小

视觉编码器与投影层以 16 位存储、以 float32 计算：PNG 输入的预处理以及送入语言模型 的向量与 Transformers 一致（每行余弦相似度高于 0.9999）。NVIDIA A40（Vulkan）上 q8_2f 生成 167 token/s，640×480 图像（300 个 token）编码约 3 s。cortiq serve 支持 OpenAI 格式的图像输入；视频用 --video（Y4M 文件或帧目录加 --video-fps）， 按每秒 2 帧采样。Apple silicon 尚未测量。这些文件是 Liquid AI LFM3-VL-2B 的修改形式，仍受 LFM Open License v1.0 约束（副本见 LICENSE）；年收入 1000 万美元及以上的实体进行商业使用不在许可范围内。