DevelopersDocumentationProductSolutionsESGPricing
Sign upGet Started
🧠 Build Your Own AI

Bring Your Own
Local LLM

CodexSigma is built for local AI from the ground up. Download, fine-tune, or build your own model — it plugs right in. No API keys. No data leaving your machine. Zero per-query cost.

Why Local AI

Your Code, Your Model, Your Rules

Every other IDE sends your code to someone else's server. CodexSigma runs on your machine, with your model.

🔒

100% Private

Your code never leaves your laptop. No telemetry, no cloud calls, no data collection. Fine for HIPAA, PCI, GDPR from day one.

💰

Zero Per-Query Cost

Once the model is downloaded, every query is free. No API bills, no token counting, no surprise charges at scale.

⚡

Works Offline

No internet needed after setup. Works on planes, in dorms, behind firewalls, in air-gapped environments.

Model Options

Pick Your Model

CodexSigma works with any model that Ollama supports — from 1.5B to 70B parameters.

🤖

DeepSeek-Coder 6.7B

Best for general coding. Strong at Python, JavaScript, TypeScript. 4GB RAM. Runs on any laptop.

ollama pull deepseek-coder:6.7b
⚡

Qwen2.5-Coder 7B

Excellent instruction following. Supports tools/function calling natively. 4.5GB RAM.

ollama pull qwen2.5-coder:7b
🧠

CodeLlama 13B

Stronger reasoning for complex tasks. 8GB RAM. Recommended for teams with decent hardware.

ollama pull codellama:13b
🚀

Llama 3.2 3B

Fastest option. Runs on any machine including Raspberry Pi. Good for quick edits and simple tasks.

ollama pull llama3.2:3b
🏆

Qwen3-Coder 30B

Best quality. Near GPT-4 level coding. Requires 20GB RAM and GPU recommended.

ollama pull qwen3-coder:latest
🧠

EngineMind

Enterprise-grade LLM for code generation and reasoning. IBM-influenced architecture. Optimized for business logic and compliance workflows.

ollama pull enginemind
export AI_MODEL=enginemind
🍎

MLX Models (Mac)

Apple Silicon optimized. Runs 2-3x faster than Ollama on Mac. Native Metal acceleration.

pip install mlx-lm && mlx_lm.generate
Fine-Tune Your Own

Custom Model for Your Codebase

Fine-tune a model on your company's code, style, and standards. Then use it in CodexSigma.

1

Prepare your dataset

Collect code examples from your repos. Format as instruction-response pairs. Use axolotl or unsloth for training.

2

Fine-tune with LoRA

Train on a single GPU. LoRA adapters are ~100MB. Keeps the base model unchanged.

python -m axolotl.cli.train config.yml
3

Export to GGUF

Convert your fine-tuned model to GGUF format for Ollama compatibility.

python convert.py --outfile model.gguf
4

Create Ollama model

Create a Modelfile and import into Ollama. CodexSigma detects it automatically.

ollama create my-custom-model -f Modelfile
export AI_MODEL=my-custom-model
Quick Start

Use Your Model in CodexSigma

Once your model is in Ollama, CodexSigma finds it automatically.

# 1. List available models ollama list # 2. Set your model export AI_MODEL=my-custom-model # 3. Start CodexSigma JWT_SECRET=$(node -e "console.log(require('crypto').randomBytes(48).toString('hex'))") \ AI_MODEL=my-custom-model \ npx tsx apps/ide/packages/server/src/index.ts # ✓ Dr. Q will use your custom model
Multi-Provider Fallback

Never Get Stuck

Configure multiple models. CodexSigma falls through automatically if one fails.

🦙

Ollama (Local)

Your custom model runs first. Zero latency. No network needed.

🤖

Groq (Cloud, Free)

Free 70B model via API. Falls back if local model is overloaded.

☁️

OpenAI/Anthropic

Final fallback. Only used when local and free options are exhausted.