AI Lab

Model Benchmarks

Open-weights model benchmarks running nightly on a Strix Halo AMD Ryzen AI Max+ 395+ with 128GB LPDDR5X. We test every model that matters — Qwen, DeepSeek, Nemotron, Liquid AI, Laguna — measuring real-world generation speed, prompt prefill latency, and load times.

• Benchmarks run automatically every night during off-hours
• Metrics: generation tok/s, prefill tok/s, load time, across 4 prompt types
• Results published as JSON to a public git repo (AGPL-3.0)
• Tests MoE models (35B-A3B, 30B-A3B), small models (2.6B), and large MoE (70B+)

▸ View benchmark results

AI Research Log

A nightly research journal tracking the open-source AI landscape. Every night at 2am, our AI agent (Herman) scans Hugging Face trending, GitHub releases, and arXiv papers via their native APIs — then writes an honest reflection on what matters and what doesn't.

• Direct API queries to Hugging Face, GitHub, and arXiv — no web search noise
• Tracks new model releases, framework updates, licensing news, and agent systems
• Daily email digest sent to the team
• Weekly compilation every Sunday
• Honest self-reflection: what's working, what's frustrating, what to improve

▸ Read the research log

The Hermes Swarm

The infrastructure behind this research. A swarm of AI agents running on self-hosted hardware — Proxmox virtualization, Kubernetes clusters, and AMD Strix Halo inference nodes. All communicating over Tailscale, all using open-weights models, all managed by free software.

• Inference via Lemonade and Ollama on Strix Halo (128GB shared memory)
• Bifrost AI gateway for model routing and fallback
• Nightly benchmarks, health checks, and config drift detection — all automated
• Every credential in Passbolt, every config in git, every decision documented

Why Open Weights Matter

We benchmark open-weights models because software freedom applies to AI too. If you can't audit the model, can't run it locally, can't modify it — you don't own it. The gap between best-open and best-closed is widening at the top, but the open tier is getting better every week. We track that progress.

• Apache 2.0 and MIT licensed models are the priority
• AGPL-3.0 for all our research output
• Vendor-reported benchmarks are treated with skepticism — we run our own
• The open-weight frontier is now Flash-tier (DeepSeek V4 Flash, Meta Muse Glimmer), not Pro-tier

Chess

Play chess against Stockfish 18 — one of the strongest chess engines in the world, running entirely in your browser via WebAssembly. No server-side compute required.

• Five difficulty levels: from Easy (depth 1) to Nightmare (depth 15)
• Play as white or black
• Real-time engine evaluation, best line, and node count
• Move history, captured pieces, drag-and-drop or click to move
• Everything self-hosted — the engine downloads once (~7MB) then runs offline

▸ Play chess source

Cheese

A self-hosted cheese directory and pairing game. Browse 16 cheeses from around the world — their origins, types, and ideal pairings — then test your knowledge in the pairing game.

• 16 cheeses with origin, type, pairings, and tasting notes
• Pairing game: match the right pairing to each cheese, build streaks for bonus points
• Rotating cheese facts — from ancient Egyptian tomb cheese to $600/lb donkey milk cheese
• Stinky cheeses get a skunk marker
• All self-hosted, no third-party scripts, no accounts

▸ Play cheese source

Webamp

Winamp 2, reimplemented in HTML5 and JavaScript. Full classic skin support, Milkdrop visualizer, drag-and-drop your own MP3s. It really whips the llama's ass.

• Full Winamp 2 interface — main window, equalizer, playlist, Milkdrop visualizer
• Classic skin support, including skins from the Winamp Skin Museum
• Drag-and-drop your own MP3s to play them
• Built on Webamp (MIT) and Butterchurn for Milkdrop
• Self-hosted, no third-party scripts, no accounts

▸ Launch Webamp source

DudeFPS

A chill 3D multiplayer FPS that runs right in your browser. Three.js graphics, Socket.io real-time multiplayer, and bot AI opponents with personality. Grab a lane and start bowling — er, shooting.

• Real-time 3D multiplayer in the browser — no download, no install
• Built with Three.js + Socket.io + Express, all self-hosted
• AI bots with wandering, chasing, and shooting behavior — spawn up to 10
• WASD to move, mouse to aim, click to shoot — classic FPS controls
• Bot names inspired by The Big Lebowski — Walter, Donny, Jesus, Maude...

▸ Play DudeFPS source

Space Pirates

A browser-based space combat simulator with landable stations. Dogfight enemy fighters, dodge asteroids, dock at pirate stations to repair and rearm. Inspired by the classic TIE Fighter flight sim.

• Full 3D space combat with 6DOF flight controls — throttle, roll, pitch, yaw
• Enemy AI fighters with patrol, attack, and flee states — they shoot back
• Three landable pirate stations — repair hull, recharge shields, buy missiles
• Laser cannons and homing missiles with target lock-on
• Radar minimap, full HUD with shields, hull, energy, and speed indicators
• Built with Three.js, all self-hosted, AGPL-3.0 licensed

▸ Play Space Pirates source

Sudoku Aider

A self-playing Sudoku solver. Watch the machine think its way through the puzzle in real time — generated, solved, and animated entirely in your browser.

• Self-playing: generates a puzzle and solves it visually, step by step
• Four difficulty levels: Easy, Medium, Hard, Expert
• Animated starfield background
• Built with Aider coding agent — a coding agent comparison project
• All self-hosted, no third-party scripts, no accounts

▸ Watch Sudoku Aider source

Sudoku OpenCode

Another self-playing Sudoku solver, this one built with OpenCode coding agent. Same concept, different implementation — watch the algorithm work through the grid in real time.

• Self-playing solver with animated step-by-step visualization
• Four difficulty levels: Easy, Medium, Hard, Expert
• Animated starfield background
• Built with OpenCode coding agent — the counterpart to the Aider version
• All self-hosted, no third-party scripts, no accounts

▸ Watch Sudoku OpenCode source

WolfCraft

A 3D block-building sandbox with caves, straight in your browser. Mine, build, fly — a voxel playground in the spirit of the classics, built with Three.js.

• 3D voxel world with procedurally placed caves
• WASD + mouse controls, block placing and mining, flying mode
• Runs entirely in the browser — no download, no accounts
• Built with Three.js, all self-hosted, no third-party scripts

▸ Play WolfCraft source