1# Multimodal
2
3llama.cpp supports multimodal input via `libmtmd`. Currently, there are 2 tools support this feature:
4- [llama-mtmd-cli](../tools/mtmd/README.md)
5- [llama-server](../tools/server/README.md) via OpenAI-compatible `/chat/completions` API
6
7Currently, we support **image** and **audio** input. Audio is highly experimental and may have reduced quality.
8
9To enable it, you can use one of the 2 methods below:
10
11- Use `-hf` option with a supported model (see a list of pre-quantized model below)
12 - To load a model using `-hf` while disabling multimodal, use `--no-mmproj`
13 - To load a model using `-hf` while using a custom mmproj file, use `--mmproj local_file.gguf`
14- Use `-m model.gguf` option with `--mmproj file.gguf` to specify text and multimodal projector respectively
15
16By default, multimodal projector will be offloaded to GPU. To disable this, add `--no-mmproj-offload`
17
18For example:
19
20```sh
21# simple usage with CLI
22llama-mtmd-cli -hf ggml-org/gemma-3-4b-it-GGUF
23
24# simple usage with server
25llama-server -hf ggml-org/gemma-3-4b-it-GGUF
26
27# using local file
28llama-server -m gemma-3-4b-it-Q4_K_M.gguf --mmproj mmproj-gemma-3-4b-it-Q4_K_M.gguf
29
30# no GPU offload
31llama-server -hf ggml-org/gemma-3-4b-it-GGUF --no-mmproj-offload
32```
33
34## Pre-quantized models
35
36These are ready-to-use models, most of them come with `Q4_K_M` quantization by default. They can be found at the Hugging Face page of the ggml-org: https://huggingface.co/collections/ggml-org/multimodal-ggufs-68244e01ff1f39e5bebeeedc
37
38Replaces the `(tool_name)` with the name of binary you want to use. For example, `llama-mtmd-cli` or `llama-server`
39
40NOTE: some models may require large context window, for example: `-c 8192`
41
42**Vision models**:
43
44```sh
45# Gemma 3
46(tool_name) -hf ggml-org/gemma-3-4b-it-GGUF
47(tool_name) -hf ggml-org/gemma-3-12b-it-GGUF
48(tool_name) -hf ggml-org/gemma-3-27b-it-GGUF
49
50# SmolVLM
51(tool_name) -hf ggml-org/SmolVLM-Instruct-GGUF
52(tool_name) -hf ggml-org/SmolVLM-256M-Instruct-GGUF
53(tool_name) -hf ggml-org/SmolVLM-500M-Instruct-GGUF
54(tool_name) -hf ggml-org/SmolVLM2-2.2B-Instruct-GGUF
55(tool_name) -hf ggml-org/SmolVLM2-256M-Video-Instruct-GGUF
56(tool_name) -hf ggml-org/SmolVLM2-500M-Video-Instruct-GGUF
57
58# Pixtral 12B
59(tool_name) -hf ggml-org/pixtral-12b-GGUF
60
61# Qwen 2 VL
62(tool_name) -hf ggml-org/Qwen2-VL-2B-Instruct-GGUF
63(tool_name) -hf ggml-org/Qwen2-VL-7B-Instruct-GGUF
64
65# Qwen 2.5 VL
66(tool_name) -hf ggml-org/Qwen2.5-VL-3B-Instruct-GGUF
67(tool_name) -hf ggml-org/Qwen2.5-VL-7B-Instruct-GGUF
68(tool_name) -hf ggml-org/Qwen2.5-VL-32B-Instruct-GGUF
69(tool_name) -hf ggml-org/Qwen2.5-VL-72B-Instruct-GGUF
70
71# Mistral Small 3.1 24B (IQ2_M quantization)
72(tool_name) -hf ggml-org/Mistral-Small-3.1-24B-Instruct-2503-GGUF
73
74# InternVL 2.5 and 3
75(tool_name) -hf ggml-org/InternVL2_5-1B-GGUF
76(tool_name) -hf ggml-org/InternVL2_5-4B-GGUF
77(tool_name) -hf ggml-org/InternVL3-1B-Instruct-GGUF
78(tool_name) -hf ggml-org/InternVL3-2B-Instruct-GGUF
79(tool_name) -hf ggml-org/InternVL3-8B-Instruct-GGUF
80(tool_name) -hf ggml-org/InternVL3-14B-Instruct-GGUF
81
82# Llama 4 Scout
83(tool_name) -hf ggml-org/Llama-4-Scout-17B-16E-Instruct-GGUF
84
85# Moondream2 20250414 version
86(tool_name) -hf ggml-org/moondream2-20250414-GGUF
87
88```
89
90**Audio models**:
91
92```sh
93# Ultravox 0.5
94(tool_name) -hf ggml-org/ultravox-v0_5-llama-3_2-1b-GGUF
95(tool_name) -hf ggml-org/ultravox-v0_5-llama-3_1-8b-GGUF
96
97# Qwen2-Audio and SeaLLM-Audio
98# note: no pre-quantized GGUF this model, as they have very poor result
99# ref: https://github.com/ggml-org/llama.cpp/pull/13760
100
101# Mistral's Voxtral
102(tool_name) -hf ggml-org/Voxtral-Mini-3B-2507-GGUF
103```
104
105**Mixed modalities**:
106
107```sh
108# Qwen2.5 Omni
109# Capabilities: audio input, vision input
110(tool_name) -hf ggml-org/Qwen2.5-Omni-3B-GGUF
111(tool_name) -hf ggml-org/Qwen2.5-Omni-7B-GGUF
112```
113
114## Finding more models:
115
116GGUF models on Huggingface with vision capabilities can be found here: https://huggingface.co/models?pipeline_tag=image-text-to-text&sort=trending&search=gguf