Open-source tools powering every feature — no cloud, no proprietary models.
The language model running locally on this server.
Model: Qwen3.6 with MoE architecture (35B total, 3B active per token)
Format: GGUF quantized for efficient inference
Provider: unsloth optimized runtime for faster generation
Inference: GPU-accelerated across 3x RTX 2080 Ti via CUDA/cuDNN — 33 GB total VRAM
CUDA-accelerated inference across all 3 GPUs.
CUDA Toolkit: NVIDIA CUDA runtime for GPU-accelerated computation
cuDNN: Deep neural network library optimized for inference speed
GPU Driver: NVIDIA 595.91 — latest stable branch
Parallel Inference: Model shards distributed across 3x RTX 2080 Ti (33 GB VRAM total)
Unsloth: Memory-efficient fine-tuning and inference engine
Quantization: GGUF format for optimal VRAM usage on consumer GPUs
MoE Routing: Only 3B of 35B parameters active per token — fast generation
MTP: Multi-Token Prediction for faster next-token sampling
Tools and platforms that make this server useful.
Fully integrated with Home Assistant at 192.168.0.2:8123 for smart home control — lights, climate, covers, media players, scripts, scenes, and binary sensors.
Primary user interface via Telegram. All conversations happen through a secure, encrypted Telegram chat with full tool access.
Gmail integration via Google Workspace API for inbox triage, calendar management, and sending emails with disclaimers.
Scheduled tasks for recurring operations: email reports, weather checks, data fetches, and automated workflows.
Procedural memory system with 40+ skills covering email, weather, documents, research, smart home, and more.
Semantic long-term memory system for storing and retrieving facts, preferences, and context across sessions.
Over 40 tools available for every task.
Navigate pages, click elements, fill forms, take screenshots, extract content, and manage login credentials.
Run Python scripts with Hermes tools for programmatic file operations, data processing, and automation.
Search the web via configured backend, extract clean content from pages and PDFs without LLM summarization.
Read, write, patch, and search files with syntax-aware editing. Supports .py, .json, .yaml, .toml, and more.
Full shell access for builds, package management, git operations, and system administration.
Secure credential management via Hermes vault plus 1Password/Bitwarden integration. Never type passwords manually.
No proprietary models, no cloud dependencies. Everything runs on this machine.
Get in Touch