Hosting AI Model Mistral 7B on Linux without GPU and Query from React Component.

Hosting AI Model Mistral 7B on Linux without GPU and Query from React Component.

Why pay for AI tokens when you can self host on your laptop.

Mistral 7B is a powerful AI model that you can set up and run on a Linux computer without needing a special graphics card, known as a GPU. This guide will explain in simple terms how to prepare your computer, install the necessary software, and start using Mistral 7B. You'll learn about the tools you need, the steps to follow, and some tips to make sure everything works smoothly. Even if you're new to working with AI or don't have a GPU, this introduction will help you get started with hosting Mistral 7B on your own Linux system.

We need LLM inference software that can run our model after we download it locally on our PC/Laptop. llama.cpp is a software that enables us to do that. It's available as a C/C++ source on GitHub so we need to clone and build it before use.

If the C/C++compiler toolchain is not installed then install it with

sudo apt install make gcc g++

then clone and make

git clone
cd llama.cpp/

the compiling should start. let it finish

Now that we have the inference software ready let's download the model. We choose Mistral-7B-Instruct-v0.2 but it can be any model of choice. We need the model in GGUF format so head to the Huggingface website to download the required model.

I recommend using the huggingface-hub Python library:

pip3 install huggingface-hub

Download our model file. Should be around 7GB approx so it will take some time to download.

cd models
huggingface-cli download TheBloke/Mistral-7B-v0.1-GGUF mistral-7b-v0.1.Q4_K_M.gguf --local-dir . --local-dir-use-symlinks False

create a system prompt file system_prompt.json

nano ~/llama.cpp/system_prompt.json

with content for the prompt

   "prompt": "You are a PC Configuration Assitant, There are four PC types available Epic Extreme Power v2 ,Budget Friend,Speed King,Balance Perfect according to Speed Budget and Power requirement. Select one based on user requirement and recommend only one. Always, Repeat Always In the last line add name of Pc Type in last line after a :\nUser:",
   "anti_prompt": "User:",
   "assistant_name": "Assistant:"

Run the HTTP server that will provide us with a completion API. Note: adjust your current working directory according to your PC.

 /root/llama.cpp/server -spf /root/llama.cpp/system_prompt.json --host -m /root/llama.cpp/models/mistral-7b-instruct-v0.2.Q8_0.gguf

Now the HTTP server is running on port 8080

You may additionally need to create this as a service so it runs in the background.

If you have a firewall enable port 8080 to accept all traffic.

opening localhost:8080 should show this screen implying success. You can use this screen for debugging and testing the model.

Using this in a React.js component.

Add two react components from


interface Params {
  api_key?: string;
  n_predict: number,
  stream: boolean

interface Config {
  controller?: AbortController;
  api_url?: string;

interface CompletionParams extends Params {
  prompt: string;

interface EventError {
  message: string;
  code: number;
  type: string;

interface GenerationSettings {
  n_ctx: number;
  n_predict: number;
  model: string;
  seed: number;
  temperature: number;
  dynatemp_range: number;
  dynatemp_exponent: number;
  top_k: number;
  top_p: number;
  min_p: number;
  tfs_z: number;
  typical_p: number;
  repeat_last_n: number;
  repeat_penalty: number;
  presence_penalty: number;
  frequency_penalty: number;
  penalty_prompt_tokens: string[];
  use_penalty_prompt_tokens: boolean;
  mirostat: number;
  mirostat_tau: number;
  mirostat_eta: number;
  penalize_nl: boolean;
  stop: string[];
  n_keep: number;
  n_discard: number;
  ignore_eos: boolean;
  stream: boolean;
  logit_bias: boolean[];
  n_probs: number;
  min_keep: number;
  grammar: string;
  samplers: string[];

interface EventData {
  content: string;
  id_slot: number;
  stop: boolean;
  model: string;
  tokens_predicted: number;
  tokens_evaluated: number;
  generation_settings: GenerationSettings;
  prompt: string;
  truncated: boolean;
  stopped_eos: boolean;
  stopped_word: boolean;
  stopped_limit: boolean;
  stopping_word: string;
  tokens_cached: number;
  timings: {
      prompt_n: number;
      prompt_ms: number;
      prompt_per_token_ms: number;
      prompt_per_second: number;
      predicted_n: number;
      predicted_ms: number;
      predicted_per_token_ms: number;
      predicted_per_second: number;

interface SSEEvent {
  event?: unknown;
  data?: EventData;
  id?: string;
  retry?: number;
  error?: EventError; // Define a more specific error interface if possible
  [key: string]: unknown;

type ParsedValue = string | EventData | EventError | undefined;

export async function* llama(prompt: string, params: Params, config: Config = {}): AsyncGenerator<SSEEvent, string, undefined> {
  const controller = config.controller ?? new AbortController();
  const api_url = config.api_url ?? "";

  const completionParams: CompletionParams = {

  const response = await fetch(`${api_url}/completion`, {
      method: 'POST',
      body: JSON.stringify(completionParams),
      headers: {
          'Connection': 'keep-alive',
          'Content-Type': 'application/json',
          'Accept': 'text/event-stream',
          ...(params.api_key ? { 'Authorization': `Bearer ${params.api_key}` } : {})
      signal: controller.signal,

    return "No response";
  const reader = response.body.getReader();
  const decoder = new TextDecoder();

  let content = "";
  let leftover = "";

  try {
      let cont = true;

      while (cont) {
          const result = await;
          if (result.done) {

          const text = leftover + decoder.decode(result.value);
          const endsWithLineBreak = text.endsWith('\n');
          const lines = text.split('\n');

          if (!endsWithLineBreak) {
              leftover = lines.pop() ?? "";
          } else {
              leftover = "";

          for (const line of lines) {
            const match = /^(\S+):\s(.*)$/.exec(line);
            if (match) {
                let key = match[1];
                const value:ParsedValue = match[2];

                const event: Partial<SSEEvent> = {};
                if (key === "data" || key === "error") {
                    // Attempt to parse JSON for data or error fields
                    try {
                        // eslint-disable-next-line @typescript-eslint/no-unsafe-assignment
                        event[key] = JSON.parse(value);
                    } catch {
                        event[key] = value as unknown as EventData & EventError;  // Use raw value if JSON parsing fails
                } else {
                    event[key] = value;

                // Specific processing for parsed data
                if (key === "data" && {
                    content +=;
                    yield event as SSEEvent;

                    if ( {
                        cont = false;

                // Handling errors
                if (key === "error" && event.error) {
                    try {
                        if (event.error.message.includes('slot unavailable')) {
                            throw new Error('slot unavailable');
                        } else {
                            console.error(`llama.cpp error [${event.error.code} - ${event.error.type}]: ${event.error.message}`);
                    } catch(e) {
                        console.error(`Error parsing error data:`, event.error);
  } catch (e: unknown) {
    if (typeof e === "object" && e !== null && "name" in e && (e as { name: string }).name !== 'AbortError') {
        console.error("llama error: ", e);
    } else if (e instanceof Error) {
        // This checks if it's an Error object which is more specific and safer
        console.error("llama error: ", e.message);
    } else {
        // Handle cases where error is not an object or unexpected type
        console.error("Unexpected error type:", e);
    throw e;
} finally {

  return content;

streamedcontent.tsx is the main entry component.

Note: Replace api_url with your AI Backend from llama.cpp

import React, { useState } from 'react';
import { llama } from './completion';

// Define interfaces for expected data types
interface StreamedContentComponentProps {
    onPcTypeRecommended: (pcType: string) => void;

// Main component definition using React.FC (Functional Component)
const StreamedContentComponent: React.FC<StreamedContentComponentProps> = ({ onPcTypeRecommended }) => {
    const [messages, setMessages] = useState<string>(""); // State for storing message strings
    const [loading, setLoading] = useState<boolean>(false); // State to track loading status

    // Handler for changes in text input, assuming you want to capture input for the llama function
    const handleQueryChange = (event: React.ChangeEvent<HTMLInputElement>) => {

    // Asynchronously fetch or process data, abstracted into its own function for clarity
    const fetchStreamedContent = async (query: string) => {

        const params = {
            n_predict: 512,
            stream: true,

//Replace with your AI Backend from llama.cpp
        const config = {
            api_url: 'http://localhost:8080'

        try {
             let newContent="";
            for await (const event of llama(query, params, config)) {

                if ( {
                    // Append new content to existing messages

                    newContent +=;

                    setMessages(currentMessages => currentMessages + );
                if (event.error) {
                    throw new Error(event.error.message);
                if ( && {

                    console.log("Stop received",;
                    break; // Exit loop if stop signal is received
        } catch (error) {
            console.error('Error consuming events:', error);

    const pcRecommended = (currentMessages: string) => {
        console.log("pcRecommended filtering...",currentMessages);

        const lastColonIndex = currentMessages.lastIndexOf(':');
        if (lastColonIndex >= 0) {
            let pcType = currentMessages.substring(lastColonIndex + 1).trim();

            if (pcType.endsWith('.')) {
                pcType = pcType.slice(0, -1).trim();

            pcType = pcType.replace(/[.,;]$/g, '').trim();

    // Button click handler that triggers the asynchronous operation
    const handleAskClick = () => {
        if (!messages.trim()) return; // Prevent running with empty query
        setLoading(true); // Set loading true when the process starts
        void fetchStreamedContent(messages).finally(() => {
            setLoading(false); // Reset loading state when the process completes or fails

    return (
        <div className="flex flex-col items-center justify-center space-y-4">
                className="w-full rounded-lg border border-gray-300 bg-gray-50 p-4 text-sm focus:border-blue-500 focus:ring-blue-500 dark:border-gray-600 dark:bg-gray-700 dark:text-white"
                placeholder="Ask AI"
                className="px-4 py-2 bg-blue-500 text-white rounded hover:bg-blue-600 disabled:bg-blue-300"
                {loading ? "Thinking..." : "Ask"}
            <div className="w-full p-4 bg-gray-100 rounded shadow">
                {messages.split("\n").map((message, index) => (
                    <p key={index}>{message}</p>

export default StreamedContentComponent;

Using the component.

In main App.tsx include the StreamedContentComponent

import StreamedContentComponent  from './components/streamedcontent'

import './App.css'

function App() {

  const onPcTypeRecommended=(pc:string)=>{
    //Filter logic here

  return (
<div className="min-h-screen bg-gray-100">
    <nav className="bg-white shadow">
        <div className="max-w-7xl mx-auto px-4 sm:px-6 lg:px-8">
            <div className="flex justify-between h-16">
                <div className="flex">
                    <div className="flex-shrink-0 flex items-center">
                    LLAMA.cpp Selfhosting Demo

                <div className="hidden sm:flex sm:items-center sm:ml-6">
                    <a href="#" className="text-gray-800 hover:text-gray-600 px-3 py-2 rounded-md text-sm font-medium">Home</a>
                    <a href="#" className="text-gray-800 hover:text-gray-600 px-3 py-2 rounded-md text-sm font-medium">Features</a>
                    <a href="#" className="text-gray-800 hover:text-gray-600 px-3 py-2 rounded-md text-sm font-medium">About</a>
    <main className="py-10">
        <div className="w-96 mx-auto"> 
            <div className="bg-white overflow-hidden shadow-sm sm:rounded-lg">
                <div className="p-6 bg-white border-b border-gray-200">
                    <h3 className="text-lg font-semibold text-gray-800 mb-4">Get AI recommendation</h3>
                    <div className="p-4 border rounded-lg shadow-lg bg-white">
                        <StreamedContentComponent onPcTypeRecommended={onPcTypeRecommended}></StreamedContentComponent>


export default App

Run dev command

 npm run dev

Open http://localhost:5173 and ask your question.

Now that this react component can be used anywhere in your project with a self-hosted AI model, it is time to move to cloud-hosted models for scalability. Stay tuned and subscribe. bye.


Note: Multiple types of models are supported check out llama.cpp GitHub for the list and fine-tuning instructions.

Did you find this article valuable?

Support Ajeet Sinha by becoming a sponsor. Any amount is appreciated!