mistral.rs Integration

Run LLMs locally with native Rust inference - no external servers, no API keys.

---

What is mistral.rs?

mistral.rs is a high-performance Rust inference engine that runs LLMs directly on your hardware. ADK-Rust integrates it through the adk-mistralrs crate.

Key highlights:

• 🦀 Native Rust - No Python, no external servers
• 🔒 Fully offline - No API keys or internet required
• ⚡ Hardware acceleration - CUDA, Metal, CPU optimizations
• 📦 Quantization - Run large models on limited hardware
• 🔧 LoRA adapters - Fine-tuned model support with hot-swapping
• 👁️ Vision models - Image understanding capabilities
• 🎯 Multi-model - Serve multiple models from one instance

---

Step 1: Add Dependencies

Since adk-mistralrs depends on git repositories, it cannot be published to crates.io. Add it via git:

[package]
name = "my-local-agent"
version = "0.1.0"
edition = "2024"

[dependencies]
adk-mistralrs = { git = "https://github.com/zavora-ai/adk-rust" }
adk-agent = { git = "https://github.com/zavora-ai/adk-rust" }
adk-rust = { git = "https://github.com/zavora-ai/adk-rust" }
tokio = { version = "1", features = ["full"] }
anyhow = "1.0"

For hardware acceleration, add feature flags:

# macOS with Apple Silicon
adk-mistralrs = { git = "https://github.com/zavora-ai/adk-rust", features = ["metal"] }

# NVIDIA GPU (requires CUDA toolkit)
adk-mistralrs = { git = "https://github.com/zavora-ai/adk-rust", features = ["cuda"] }

---

Step 2: Basic Example

Load a model from HuggingFace and run it locally:

use adk_agent::LlmAgentBuilder;
use adk_mistralrs::{Llm, MistralRsConfig, MistralRsModel, ModelSource};
use adk_rust::Launcher;
use std::sync::Arc;

#[tokio::main]
async fn main() -> anyhow::Result<()> {
    // Load model from HuggingFace (downloads on first run)
    let config = MistralRsConfig::builder()
        .model_source(ModelSource::huggingface("microsoft/Phi-3.5-mini-instruct"))
        .build();

    println!("Loading model (this may take a while on first run)...");
    let model = MistralRsModel::new(config).await?;
    println!("Model loaded: {}", model.name());

    // Create agent
    let agent = LlmAgentBuilder::new("local_assistant")
        .description("Local AI assistant powered by mistral.rs")
        .instruction("You are a helpful assistant running locally. Be concise.")
        .model(Arc::new(model))
        .build()?;

    // Run interactive chat
    Launcher::new(Arc::new(agent)).run().await?;

    Ok(())
}

What happens:

1. First run downloads the model from HuggingFace (~2-8GB depending on model)
2. Model is cached locally in ~/.cache/huggingface/
3. Subsequent runs load from cache instantly

---

Step 3: Reduce Memory with Quantization

Large models need lots of RAM. Use ISQ (In-Situ Quantization) to reduce memory:

use adk_agent::LlmAgentBuilder;
use adk_mistralrs::{Llm, MistralRsConfig, MistralRsModel, ModelSource, QuantizationLevel};
use adk_rust::Launcher;
use std::sync::Arc;

#[tokio::main]
async fn main() -> anyhow::Result<()> {
    // Load model with 4-bit quantization for reduced memory
    let config = MistralRsConfig::builder()
        .model_source(ModelSource::huggingface("microsoft/Phi-3.5-mini-instruct"))
        .isq(QuantizationLevel::Q4_0) // 4-bit quantization
        .paged_attention(true) // Memory-efficient attention
        .build();

    println!("Loading quantized model...");
    let model = MistralRsModel::new(config).await?;
    println!("Model loaded: {}", model.name());

    let agent = LlmAgentBuilder::new("quantized_assistant")
        .instruction("You are a helpful assistant. Be concise.")
        .model(Arc::new(model))
        .build()?;

    Launcher::new(Arc::new(agent)).run().await?;

    Ok(())
}

Quantization levels:

---

Step 4: LoRA Adapters (Fine-Tuned Models)

Load models with LoRA adapters for specialized tasks:

use adk_agent::LlmAgentBuilder;
use adk_mistralrs::{AdapterConfig, Llm, MistralRsAdapterModel, MistralRsConfig, ModelSource};
use adk_rust::Launcher;
use std::sync::Arc;

#[tokio::main]
async fn main() -> anyhow::Result<()> {
    // Load base model with LoRA adapter
    let config = MistralRsConfig::builder()
        .model_source(ModelSource::huggingface("meta-llama/Llama-3.2-3B-Instruct"))
        .adapter(AdapterConfig::lora("username/my-lora-adapter"))
        .build();

    println!("Loading model with LoRA adapter...");
    let model = MistralRsAdapterModel::new(config).await?;
    println!("Model loaded: {}", model.name());
    println!("Available adapters: {:?}", model.available_adapters());

    let agent = LlmAgentBuilder::new("lora_assistant")
        .instruction("You are a helpful assistant with specialized knowledge.")
        .model(Arc::new(model))
        .build()?;

    Launcher::new(Arc::new(agent)).run().await?;

    Ok(())
}

Hot-swap adapters at runtime:

model.swap_adapter("another-adapter").await?;

---

Step 5: Vision Models (Image Understanding)

Process images with vision-language models:

use adk_mistralrs::{Llm, MistralRsConfig, MistralRsVisionModel, ModelSource};

#[tokio::main]
async fn main() -> anyhow::Result<()> {
    let config = MistralRsConfig::builder()
        .model_source(ModelSource::huggingface("microsoft/Phi-3.5-vision-instruct"))
        .build();

    println!("Loading vision model...");
    let model = MistralRsVisionModel::new(config).await?;
    println!("Model loaded: {}", model.name());

    // Analyze an image
    let image = image::open("photo.jpg")?;
    let response = model.generate_with_image("Describe this image.", vec![image]).await?;

    Ok(())
}

---

Step 6: Multi-Model Serving

Serve multiple models from a single instance:

use adk_mistralrs::{MistralRsConfig, MistralRsMultiModel, ModelSource};

#[tokio::main]
async fn main() -> anyhow::Result<()> {
    let multi = MistralRsMultiModel::new();

    // Add models
    let phi_config = MistralRsConfig::builder()
        .model_source(ModelSource::huggingface("microsoft/Phi-3.5-mini-instruct"))
        .build();
    multi.add_model("phi", phi_config).await?;

    let gemma_config = MistralRsConfig::builder()
        .model_source(ModelSource::huggingface("google/gemma-2-2b-it"))
        .build();
    multi.add_model("gemma", gemma_config).await?;

    // Set default and route requests
    multi.set_default("phi").await?;
    println!("Available models: {:?}", multi.model_names().await);

    // Route to specific model
    // multi.generate_with_model(Some("gemma"), request, false).await?;

    Ok(())
}

---

Model Sources

HuggingFace Hub (Default)

ModelSource::huggingface("microsoft/Phi-3.5-mini-instruct")

Local Directory

ModelSource::local("/path/to/model")

Pre-Quantized GGUF

ModelSource::gguf("/path/to/model.Q4_K_M.gguf")

---

Recommended Models

---

Hardware Acceleration

macOS (Apple Silicon)

adk-mistralrs = { git = "https://github.com/zavora-ai/adk-rust", features = ["metal"] }

Metal acceleration is automatic on M1/M2/M3 Macs.

NVIDIA GPU

adk-mistralrs = { git = "https://github.com/zavora-ai/adk-rust", features = ["cuda"] }

Requires CUDA toolkit 11.8+.

CPU Only

No features needed - CPU is the default.

---

Run Examples

# Basic usage
cargo run --bin basic

# With quantization
cargo run --bin quantized

# LoRA adapters
cargo run --bin lora

# Multi-model setup
cargo run --bin multimodel

# Vision models
cargo run --bin vision

---

Troubleshooting

Out of Memory

// Enable quantization
.isq(QuantizationLevel::Q4_0)
// Enable paged attention
.paged_attention(true)

Slow First Load

• First run downloads the model (~2-8GB)
• Subsequent runs use cached model

Model Not Found

• Check HuggingFace model ID is correct
• Ensure internet connection for first download

---

Related

• Model Providers - Cloud LLM providers
• Ollama - Alternative local model server
• LlmAgent - Using models with agents

---

Previous: ← Ollama (Local) | Next: Function Tools →