π§ Neural OS Hermes β HW Expert v3
The first AI-native bare-metal operating system.
61.453 hardware IDs recognized. 0 errors. No Linux. No POSIX.
π₯ What is Neural OS?
Neural OS Hermes is a complete bare-metal operating system written in Rust (no_std), where everything is an Agent or a Skill:
- π§ 247+ agents β every hardware driver, daemon, and service is an agent
- π HW Expert v3 β neural network that identifies 61.453 unique PCI/USB/ACPI devices
- ποΈ Self-healing β detects missing firmware, downloads it via HTTP, loads it hot. No reboot.
- π¨ 3-layer visual β Orb (FFT-reactive) + Hermes CLI overlay + Window Manager
- π€ Full audio pipeline β HDA capture + playback, FFT visualization, TTS (Piper VITS)
- π‘ WiFi β Intel AX200/AX210 with real iwlwifi ucode loading
- π‘οΈ Zero-trust β SafetyAgent with 4 invariants, fail-closed by default
ποΈ Architecture
βββββββββββββββββββββββββββββββ
β Hermes (Intent Router) β
β ReAct Β· Workflow Β· SDD Β· MCP β
ββββββββββββ¬βββββββββββββββββββ
β
ββββββββββββββββββΌβββββββββββββββββ
βΌ βΌ βΌ
ββββββββββββββ ββββββββββββββ ββββββββββββββ
β Cortex β βHW Expert v3β β Trinity β
β LLM (BitNet)β β 61K HWIDs β β MoE Routerβ
β 850M paramsβ β 1M params β β 6 experts β
ββββββββββββββ ββββββββββββββ ββββββββββββββ
β β β
βΌ βΌ βΌ
βββββββββββββββββββββββββββββββββββββββββββββββ
β 247+ Agents (hardware + system) β
β HwDetect Β· NetAgent Β· WifiAgent Β· Display β
β GpuDriver Β· HDAudio Β· UvcCam Β· SelfHeal β
β Cron Β· Security Β· Safety Β· Optimizer Β· ... β
βββββββββββββββββββββββββββββββββββββββββββββββ
β β β
βΌ βΌ βΌ
βββββββββββββββββββββββββββββββββββββββββββββββ
β Hardware Β· Firmware (116 blobs) Β· ATA β
β NVIDIA GP108 Β· Intel i915 Β· Realtek Β· iwl β
β e1000 Β· HDA Β· xHCI Β· AHCI Β· NVMe Β· TPM β
βββββββββββββββββββββββββββββββββββββββββββββββ
π― HW Expert v3 β What it does
The HW Expert is a BitNet b1.58 ternary neural network (1M params, 259KB) that runs directly in the kernel (no_std Rust). It identifies hardware devices by their PCI/USB vendor and device IDs:
| Capability | Value |
|---|---|
| Unique VID/DID recognized | 61.453 |
| Training data sources | SDIO (171K HWIDs) + pci.ids (23.8K) + usb.ids (23.9K) + kernel PCI tables (494) |
| Model architecture | 128 hidden, 6 layers, 8 heads, FFN 256 |
| Model size | 259 KB (fits in L2 cache) |
| Inference | ADD/SUB only (no FPU), 2-bit packing, 4 weights/byte |
| Loss | 3.55 β 0.389 (converged) |
When a new device is plugged in, the kernel:
- Scans PCI/USB bus β gets VID:DID
- HW Expert v3 identifies the exact device name
- SelfHealAgent checks if firmware is needed
- If missing β downloads from linux-firmware.git via HTTP β loads hot
- If no driver β LLM generates a skill β SkillObserver registers it
π Training Data
This model was trained on the largest public hardware ID collection ever assembled:
| Dataset | Records | Source |
|---|---|---|
| SDIO DriverPacks | 171.003 HWIDs | 65 Windows DriverPacks, 20K .inf files |
| pci.ids | 23.888 (2.506 vendors, 21.382 devices) | Official PCI SIG list |
| usb.ids | 23.964 (3.427 vendors, 20.537 devices) | Official USB-IF list |
| Linux kernel | 494 PCI entries | Extracted from kernel drivers/ |
| linux-firmware WHENCE | 998 entries | Official firmware manifest |
| AMD microcode | 64 patches | Family/Model/Stepping from linux-firmware |
| Total | ~220.620 |
π Getting Started
# Clone the OS
git clone https://github.com/msrovani/neural-os-core.git
cd neural-os-core
# Build (requires Rust nightly)
cargo build --release
# Run in QEMU
python tools/build_image.py
./run-qemu-uefi.ps1
# Or boot on real hardware (x86_64 UEFI)
# Use tools/disk_hw.raw as a USB drive image
π¦ What's on HuggingFace
| Artifact | Description | Link |
|---|---|---|
π§ neural-os/hw-expert-v3 |
HW identification model (this) | Model Card β |
πΎ neural-os/sdio-hwids |
171K HWIDs from 65 DriverPacks | Dataset β |
πΎ neural-os/pci-usb-ids |
48K official PCI/USB IDs | Dataset β |
πΎ neural-os/firmware-metadata |
1.2K firmware entries from WHENCE | Dataset β |
π‘ neural-os/regulatory-db |
WiFi regulations for 174 countries | Dataset β |
π License
MIT License β model and training code are open source.
Firmware blobs are MIT license from linux-firmware.git.
Neural OS Hermes β "The hardware real nΓ£o perdoa. O silΓcio obedece."
β GitHub
Inference Providers NEW
This model isn't deployed by any Inference Provider. π Ask for provider support