LocateAnything-3B is NVIDIA's 3B visual grounding model — it locates and identifies objects in images given natural language descriptions. Designed for on-device deployment and robotics applications.

These GGUFs use importance matrix (iMatrix) calibration on 2M tokens of wikitext-103: iMatrix runs calibration text through the model, measures which weights activate most, and protects them during quantization. Result: noticeably better coherence at Q2/Q3/Q4 — same file size, better output.

Quick Start

llama.cpp

llama-cli -hf liodon-ai/LocateAnything-3B-imatrix-GGUF:Q4_K_M

LM Studio / Jan

Search liodon-ai/LocateAnything-3B-imatrix-GGUF and pick your quant.

Available Quants

Quant	Size	VRAM	Notes
`IQ2_M`	1.28 GB	2 GB	ultra-tiny + iMatrix — better than standard Q2
`IQ3_M`	1.62 GB	2.5 GB	tiny + iMatrix — sharper than standard Q3
`IQ4_XS`	1.91 GB	3 GB	small + iMatrix — rivals Q5 at Q4 size
`Q2_K`	1.38 GB	2 GB	tiniest standard — runs almost anywhere, iMatrix-improved
`Q3_K_M`	1.73 GB	2.5 GB	great for 4GB VRAM, iMatrix-improved
`Q4_K_M`	2.11 GB	3 GB	sweet spot (recommended), iMatrix-improved
`Q5_K_M`	2.44 GB	4 GB	high quality, iMatrix-improved
`Q6_K`	2.80 GB	4 GB	near-lossless, iMatrix-improved
`Q8_0`	3.62 GB	5 GB	basically full quality

What is iMatrix?

Standard quantization rounds all weights equally. iMatrix:

Runs calibration text through the full-precision model
Measures which weights activate most (the "importance matrix")
Allocates more precision to important weights, less to unimportant ones

Same file size. Better output. Most noticeable at Q2/Q3/Q4.

Calibration

Importance matrix computed from 2M tokens of wikitext-103 — 128 calibration chunks.

Source Model

Original: nvidia/LocateAnything-3B
Architecture: 3B visual grounding model
Strengths: Object localization, visual grounding, on-device robotics
License: NVIDIA Open Model License