muLLM

Local-first LLM routing. Ask anything — pay almost nothing. From a CSS question to a 3D game in one conversation.

Act 1 — The Chat

Ask anything. Get an answer in under 1ms.

Groundtruth LUT answers 1,313+ common questions deterministically. No inference, no cost, no latency.

Hey, what's the difference between padding and margin in CSS?

Margin is space outside the element's border — it separates it from neighbors. Padding is space inside the border — it separates content from the edge. Margins collapse between siblings; padding never does.

⚡ 0.7ms · groundtruth · $0.00

Write me a 3D space shooter game

Routing to local 30B model… generating 847 lines of Three.js...

⚡ 4.2s · local 9B · $0.00

Act 2 — The Router

Intelligent routing. 91.9% free.

Cloud only for genuinely hard problems. Solve complex tasks in parallel.

INPUT

DeBERTa

⚡ Groundtruth <1ms

💾 Cache ~14ms

🏠 Local 9B ~5s

☁ Cloud ~3s

GT 13% · Cache 34% · Local 46% · Cloud 7%

Act 3 — The Games

263 playable games. All generated by a local model.

Code Mode routes complex generation to the 30B model. 100% HumanEval pass@1. Zero API cost for code generation.

🎯

Axe Throw Lab

Skill lab — throw, tune aim, and chase precision targets

▶ Play

💎

Bamboo Village

Village sim — small world exploration and resource play

▶ Play

🐍

Boardwalk Midway

Arcade boardwalk — compact carnival games and score loops

▶ Play

🔮

Card Toss

Throwing game — tune force and angle for clean shots

▶ Play

🤖

Crokinole 2

Tabletop skill — flick discs and control the board

▶ Play

⬆️

Foot Golf

Sports puzzle — kick through compact courses

▶ Play

Act 4 — The Models

245 3D models. Steampunk, sci-fi, fantasy.

Generated via Meshy AI pipeline, decimated and optimized. Full PBR textures. Ready for Three.js or Unity.

Steampunk Airship

View in Studio →

Iron Gear Assembly

View in Studio →

Crystal Cavern

View in Studio →

Spider Bot MkII

View in Studio →

Siege Cannon

View in Studio →

Arcane Staff

View in Studio →

Act 5 — The Proof

Self-proving. The model built its own benchmark.

muLLM routed the HumanEval runner implementation to itself, fixed the check() invocation bug, and ran all 164 problems to completion.

"The local model wrote the test harness that proved it scores 100%."

100%

HumanEval pass@1
164/164 problems

82.4%

DeBERTa accuracy
5,161 examples

65%

GameDevBench
vs GPT-4o baseline

$0.00 total API cost for 164 HumanEval solutions.

The same local 9B model that handles chat, routes queries, and evaluates results also solves algorithmic programming problems at state-of-the-art accuracy — without billing a single cent.

Act 6 — Real Savings

10,301 queries. $10.86 spent.

741 queries/day average. 13.9-day production window. Scoring log rotates at 10K — lifetime totals are higher.

queries served · $10.86 actual spend · $6,329 estimated lifetime savings vs Opus 4

91.9% Free 8.1% paid

741/day

Query avg

$0.001

Cost per query

13.9 days

Production window

muLLM

Ask anything. Get an answer in under 1ms.

Intelligent routing. 91.9% free.

263 playable games. All generated by a local model.

245 3D models. Steampunk, sci-fi, fantasy.

Self-proving. The model built its own benchmark.

10,301 queries. $10.86 spent.

From CSS question to 3D game in one conversation.