LFM2-VL-450M by Liquid AI, quantized to GGUF format for llama.cpp. Optimized for low-latency on-device vision inference, packaged for use with the RunAnywhere SDK.

Files:

LFM2-VL-450M-Q4_0.gguf — Language model Q4_0 (~209 MB)
LFM2-VL-450M-Q8_0.gguf — Language model Q8_0 (~361 MB)
mmproj-LFM2-VL-450M-Q8_0.gguf — Vision encoder (~99 MB)

Usage with RunAnywhere SDK

Swift (iOS / macOS)

swift

import RunAnywhere

RunAnywhere.registerModel(
    id: "lfm2-vl-450m-q4_0",
    name: "LFM2-VL 450M Q4_0",
    repo: "runanywhere/LFM2-VL-450M-GGUF",
    files: ["LFM2-VL-450M-Q4_0.gguf", "mmproj-LFM2-VL-450M-Q8_0.gguf"],
    framework: .llamaCpp,
    modality: .multimodal,
    memoryRequirement: 500_000_000
)

// Low-latency VLM inference
let result = try await RunAnywhere.generateVLM(
    prompt: "What do you see?",
    image: imageData,
    modelId: "lfm2-vl-450m-q4_0"
)

Kotlin (Android / JVM)

kotlin

import com.runanywhere.sdk.RunAnywhere
import com.runanywhere.sdk.models.*

RunAnywhere.registerModel(
    id = "lfm2-vl-450m-q4_0",
    name = "LFM2-VL 450M Q4_0",
    repo = "runanywhere/LFM2-VL-450M-GGUF",
    files = listOf("LFM2-VL-450M-Q4_0.gguf", "mmproj-LFM2-VL-450M-Q8_0.gguf"),
    framework = InferenceFramework.LLAMA_CPP,
    modality = ModelCategory.MULTIMODAL,
    memoryRequirement = 500_000_000L
)

val result = RunAnywhere.generateVLM(
    prompt = "What do you see?",
    image = imageData,
    modelId = "lfm2-vl-450m-q4_0"
)

Web (TypeScript)

typescript

import { RunAnywhere, LLMFramework, ModelCategory } from '@anthropic/runanywhere-web';

RunAnywhere.registerModels([{
  id: 'lfm2-vl-450m-q4_0',
  name: 'LFM2-VL 450M Q4_0',
  repo: 'runanywhere/LFM2-VL-450M-GGUF',
  files: ['LFM2-VL-450M-Q4_0.gguf', 'mmproj-LFM2-VL-450M-Q8_0.gguf'],
  framework: LLMFramework.LlamaCpp,
  modality: ModelCategory.Multimodal,
  memoryRequirement: 500_000_000,
}]);

await RunAnywhere.downloadModel('lfm2-vl-450m-q4_0');
await RunAnywhere.loadModel('lfm2-vl-450m-q4_0');

const result = await RunAnywhere.generateVLM('What do you see?', imageData, 'lfm2-vl-450m-q4_0');

React Native (TypeScript)

typescript

import { RunAnywhere } from 'runanywhere-react-native';

RunAnywhere.registerModel({
  id: 'lfm2-vl-450m-q4_0',
  name: 'LFM2-VL 450M Q4_0',
  repo: 'runanywhere/LFM2-VL-450M-GGUF',
  files: ['LFM2-VL-450M-Q4_0.gguf', 'mmproj-LFM2-VL-450M-Q8_0.gguf'],
  framework: 'llamaCpp',
  modality: 'multimodal',
  memoryRequirement: 500_000_000,
});

const result = await RunAnywhere.generateVLM('What do you see?', imageData, 'lfm2-vl-450m-q4_0');

Flutter (Dart)

dart

import 'package:runanywhere_flutter/runanywhere_flutter.dart';

RunAnywhere.registerModel(
  id: 'lfm2-vl-450m-q4_0',
  name: 'LFM2-VL 450M Q4_0',
  repo: 'runanywhere/LFM2-VL-450M-GGUF',
  files: ['LFM2-VL-450M-Q4_0.gguf', 'mmproj-LFM2-VL-450M-Q8_0.gguf'],
  framework: InferenceFramework.llamaCpp,
  modality: ModelCategory.multimodal,
  memoryRequirement: 500000000,
);

final result = await RunAnywhere.generateVLM('What do you see?', imageData, 'lfm2-vl-450m-q4_0');

Model Details

Property	Value
Base Model	LFM2-VL-450M (Liquid AI)
Parameters	450M
Quantizations	Q4_0 (~209 MB), Q8_0 (~361 MB)
Vision Encoder	mmproj Q8_0 (~99 MB)
Runtime	llama.cpp (with multimodal/mtmd)
Optimized For	Low-latency edge inference

Attribution

Original model by Liquid AI. GGUF conversion by Liquid AI.