What GPU do you need to run LFM3-VL-2B?

1.92B parameters (Hugging Face, accessed 2026-10-27). Q4_K_M GGUF: 1.1 GB (source). Architecture config (Hugging Face config.json; text_config layer_types: 6 full_attention + 5 sliding_attention (256) + 13 conv layers. The conv layers keep a small fixed-size state and no KV (arch hybrid, so the speed estimate adds the recurrent-state cost); llama.cpp lfm3 puts the attention layers in its iSWA cache (src/models/lfm3.cpp, llama-memory-hybrid-iswa). The vision encoder adds no KV cache): 6 attention layers (the rest are recurrent) × 4 KV heads × 128 head dim, plus 5 sliding-window layers (256-token window), so about 12 KB of FP16 KV cache per token of context (how we estimate context).

Liquid AI vision-language model (text, image and video in). 1.92B is Hugging Face's safetensors total of 1,918,181,888; the card splits it into a 1.73B language model, a 164M vision encoder and a 25M projector. The language model is the same 24-layer LFM3 layout as LFM3-1.7B (13 short convolutions, 5 sliding-window, 6 global attention layers) with a 128K context. Q4_K_M is Liquid AI's official GGUF (LFM3-VL-2B-Q4_K_M.gguf, 1,099,627,808 bytes, language model only); every size is in the table below. Fit ignores the vision projector (mmproj-F16 378,996,032 bytes, mmproj-Q8_0 203,031,872 bytes), which is extra if you send images. llama.cpp support (lfm3 architecture) was merged on Oct 27, 2026 (ggml-org/llama.cpp PR #31252) and is in tagged releases v0.9.0 (Nov 4, 2026) and v1.0.0 (Nov 13, 2026). The GGUF card says llama.cpp has no video input for now. Liquid's MLX builds wait on an MLX-VLM release with LFM3-VL support (in progress, per the card). LFM Open License v1.0: commercial use is licensed only while you (with your affiliates) stay under $10M annual revenue; above that, commercial use is not licensed under it. The cap does not apply to qualifying non-profits doing non-commercial or research work. Release date is Liquid's announcement; the Hugging Face repo was created Oct 19.