🇬🇧 English • 🇮🇹 Italiano • ⚡ Quick Start • 🏛️ Architecture • 🧩 Modules • 📦 GitHub
chat_record.mp4
Run local models, connect agents, use MCP tools, orchestrate workflows and work across your hardware seamlessly.
chat_record.mp4)
👉 Click here to play the full demo directly in the GitHub Video Player
Sigma Studio is a modular AI workspace that turns local hardware into an extensible AI development environment.
Within a single unified desktop interface, without wrestling with complex terminal scripts or environment conflicts, you can:
- 📥 Download any open-source model: Search, fetch, and organize models from Hugging Face or GGUF repositories with resumable multi-stream downloads.
- 💬 Chat with low-latency streaming: Converse with local models (via native SigmaEngine) or external cloud providers (OpenAI, Claude, Gemini, DeepSeek) with real-time token rendering.
- 🎭 Assign specialized roles: Switch between 20 predefined Modelfiles with 1 click (Software Architect, Coder, Mathematician, Medical Specialist, Jurist, Security Auditor...).
- 🧪 Test & benchmark hardware: Measure real tokens per second, Time-to-First-Token (TTFT), and VRAM saturation under realistic workloads.
- 🧠 Train & fine-tune: Fine-tune Small Language Models (SLMs) locally on your own GPU using Unsloth QLoRA, PEFT, and the Gradus Functional Weight Engine.
- ⚙️ Quantize locally (GGUF Forge): Convert raw FP16/FP32 weights into Q4_K_M, Q5_K_M, or Q8_0 formats directly in memory to match your hardware VRAM.
- 🔌 Equip models with system tools (MCP): Connect 12 Model Context Protocol servers to browse the web, execute terminal scripts, manage calendar/email, and control smart home devices within a watertight sandbox.
All completely free, private, and sovereign — running locally on your terms without recurring subscriptions.
Sigma Studio is engineered around a clean architectural separation between the workspace orchestrator and the underlying native inference engine:
| Component | Responsibility |
|---|---|
| Σ-SIGMA STUDIO | AI workspace and orchestration environment built around the watertight Sigma Kernel. Provides the React 19 UI, streaming multi-agent chat, MCP tool execution governance, AST sandbox security, and dynamic module loading. |
| ⚡ SIGMAENGINE | High-performance local inference engine for heterogeneous hardware. Features zero-bottleneck C++/PyTorch layer sharding across multi-GPU CUDA, CPU and system RAM offloading, Apple Metal, sub-100ms TTFT FlashAttention-2, and an in-memory GGUF Quantization Forge (Q4/Q5/Q8). |
| Problem in Local AI | Sigma Studio Solution |
|---|---|
| Local models are fragmented | Unified model & provider abstraction layer (seamlessly route between local weights and OpenAI, Claude, Gemini, DeepSeek, Groq, Ollama). |
| Large models exceed single GPU VRAM | Zero-bottleneck layer sharding & RAM offloading across multiple NVIDIA GPUs, Apple Silicon Metal, or system RAM. |
| Agents lack real system tools | Native Model Context Protocol (MCP) with 12 built-in servers (Terminal CLI, Web, Email, IoT, Memory Graph). |
| AI workflows are hard to inspect | Visual execution DAG & real-time telemetry, tracking token streaming, VRAM allocations, and tool call confirmations. |
| Extensions become monolithic bloat | Decoupled Modular Labs that can be installed on-demand from SigmaStudio-Moduli without restarting the kernel. |
| Hardware setups vary widely | Hardware-aware execution optimizing automatically for multi-GPU workstations, laptops, or edge devices like Raspberry Pi 5. |
| Local AI lacks an integrated environment | All-in-one sovereign AI development workspace: Chat, Forge, Fine-Tuning, 3D/2D Generation, Voice, and Task Automation. |
- 💻 Developers: Build AI applications, orchestrate multi-agent swarms, debug MCP tool servers, and run sandboxed code safely.
- 🔬 Researchers: Benchmark open-source LLMs, experiment with Unsloth QLoRA fine-tuning, and test distributed model layer partitioning.
- ⚡ AI Enthusiasts: Run private, sovereign frontier models locally with zero subscription fees and 100% data sovereignty.
- 🛠️ Hardware Builders: Combine heterogeneous GPUs, system RAM, and edge devices into a unified, high-throughput AI runtime.
- Workspace UI Layer: GPU-accelerated React 19 + Vite 8 frontend featuring streaming agent terminals, D3 relational memory graphs, Three.js 3D viewport, and real-time hardware telemetry.
- Sigma Kernel: Lightweight Python 3.10+ FastAPI microkernel providing strict path whitelisting, AST static analysis sandboxing, intent classification, and session management.
- Pillars of Orchestration:
- Autonomous Agents: 20 standardized Modelfiles (Architect, Developer, Mathematician, Medical Specialist, Jurist, Security Auditor, etc.).
- Providers Hub: 100% interoperable routing between local inference and cloud APIs (OpenAI, Anthropic Claude, Google Gemini, DeepSeek, Groq, Ollama).
- 12 MCP Servers: Model Context Protocol servers for filesystem, live web search, messaging, calendar, IoT, and VRAM management.
- SigmaEngine: The raw C++/PyTorch execution engine sharding layers across NVIDIA CUDA GPUs, system RAM, or Apple Metal with FlashAttention-2.
- Modular Ecosystem: Hot-loaded on demand from the community catalog
SigmaStudio-Moduli.
- ⚡ SigmaEngine Local Inference: Multi-GPU layer sharding, sub-100ms TTFT, continuous KV-cache streaming.
- 🛠️ Hugging Face Downloader & GGUF Forge: In-memory converter and quantizer (Q4_K_M, Q5_K_M, Q8_0, FP16) to fit any model to your hardware without external CLI tools.
- 🤖 Autonomous Agent Swarm & 20 Modelfiles: Persona contracts and reasoning workflows tailored for specific professional domains.
- 🔌 12 Model Context Protocol (MCP) Servers: Interactive permission governance and tool execution for system-level operations.
- 🛡️ Watertight Sandboxed Execution: Confines filesystem writes to authorized directories (
data/,scratch/,core/) with AST code protection.
- 🎨 Creative Lab 3D/2D: FLUX/SDXL text-to-image, SAM2 background removal, Hunyuan3D/TripoSR mesh generation, PBR materials.
- 🎙️ Voice Studio & Speech: Kokoro 82M ultra-fast TTS (<80ms), Coqui XTTS-v2 zero-shot voice cloning, pitch/speed tuning, live waveform visualizer.
- 🧠 Training Lab & SLM: Unsloth QLoRA, PEFT, Gradus Functional Weight Engine (FWE), Autopilot hyperparameter search.
- 🔬 Pipelines Lab & Swarm: Visual DAG pipeline designer, multi-agent research loops, step-by-step execution inspector.
- 📊 Argomenti & Knowledge Graph: D3 force-directed relational memory graph with vector RAG search.
- ⚡ Hardware & GPU Telemetry: Real-time VRAM allocation, CUDA process monitor, zombie task termination, one-click VRAM flush.
- 🏠 Smart Home Domotica: Home Assistant WebSocket/REST bridge, device control, automation triggers, climate and solar modulation.
Sigma Studio is engineered to keep the core runtime ultra-lightweight. Optional features and specialized lab environments are distributed as independently installable modules:
All modules can be installed with a single click directly inside the Hub Skills & Extensions tab in the Sigma Studio UI without restarting the server.
Zero manual configuration required: dependencies, virtual environments, hardware detection, and frontend assets are automatically verified and installed upon first launch.
git clone https://github.com/Sigmanih/SigmaStudio.git
cd SigmaStudio- Windows:
.\sigma_studio.bat - Linux / macOS / Raspberry Pi:
chmod +x sigma_studio.sh ./sigma_studio.sh
💡 On the very first run, Sigma Studio automatically creates the virtual environment, installs Python requirements, sets up the native inference runtime, builds the frontend if needed, and opens
http://localhost:8000.
- Force reinstall/update dependencies:
.\sigma_studio.bat --install(or./sigma_studio.sh --install) - Run pre-flight environment check:
.\sigma_studio.bat --check(or./sigma_studio.sh --check) - Inspect detected hardware and accelerators:
python sigma_launcher.py --info
Sigma Studio is engineered to run seamlessly across heterogeneous architectures — from multi-GPU workstation clusters to single-board edge computers:
| Platform | Architecture | Accelerators & Compute | Recommended Models | One-Click Launcher |
|---|---|---|---|---|
| Windows Workstation | x86_64 (Windows 10/11) |
NVIDIA CUDA (RTX 30xx/40xx), Vulkan, CPU | All sizes (0.5B – 70B+) | .\sigma_studio.bat |
| Linux Server & Desktop | x86_64 (Ubuntu / Debian / Arch) |
NVIDIA Multi-GPU, AMD ROCm, Intel SYCL, CPU | All sizes (0.5B – 70B+) | ./sigma_studio.sh |
| Apple Silicon Mac | arm64 (M1 / M2 / M3 / M4) |
Apple Metal (Unified Memory up to 128GB+) | 0.5B – 32B+ Q4_K_M | ./sigma_studio.sh |
| Raspberry Pi 5 / 4 | aarch64 (Debian Bookworm 64-bit) |
Broadcom Quad-Core ARM Cortex-A76 (NEON) | 0.5B – 3B SLMs (Qwen 2.5, Llama 3.2, SmolLM2) | ./sigma_studio.sh |
| Edge & PC CPU-Only | x86_64 / arm64 |
Intel / AMD AVX2/AVX-512, Snapdragon X | 0.5B – 7B Quantized (Q4_K_M) | .\sigma_studio.bat / ./sigma_studio.sh |
🍓 Raspberry Pi 5 Ready: On aarch64 Linux,
./sigma_studio.shautomatically detects the ARM Cortex CPU, selects the lightweight CPU wheel set, configures the native ARM runtime, and runs modern SLMs (Small Language Models) with zero manual setup.
Run the comprehensive Pytest kernel test suite:
pytest tests/ -vAll kernel tests validate MCP governance, agent routing, FastAPI endpoints, security sandboxing, and chat streaming with a 100% success rate.
Sigma Studio is developed as an independent, sovereign open-source project. If you find it useful for your research, workflows, or homelab setup, consider supporting continued development:
Your sponsorship supports multi-GPU inference optimizations, open-source model roles, and zero-cost sovereign AI tools.
Sigma Studio is open source software dual-licensed under:
- GNU Affero General Public License v3 (AGPL-3.0) for the open source community, developers, and researchers.
- Commercial License for enterprise deployments, proprietary integrations, and closed-source SaaS offerings.
See the LICENSE file for complete licensing terms, trademark guidelines, and commercial licensing contacts.
- CONTRIBUTING.md — Contribution guidelines, Developer Certificate of Origin, and CLA terms
- SECURITY.md — Vulnerability reporting and security policy
- CODE_OF_CONDUCT.md — Contributor Covenant v2.1
- GitHub Repository — Official repository and updates