Twin 01
Mistral 7B
Fast local reasoning twin with grouped-query attention, Swift visualizer mode, and 32 transformer blocks.
TokEmbedGQAHead
Open Live Terminal
SnapKitty Inverted NoShow Turbo3X
LLM Twin Arena
Every model gets a twin. Every twin gets a page.
Twin 02
Llama 3.2
Long-context local runtime twin for Swift challenge building, trimming, and evidence retention.
BPERoPEGQAHead
Open Live Terminal
Twin 03
Phi-3 Mini
Compact offline twin for Swift-built small proofs, local judging, and sealed challenge loops.
TokEmbedMHAHead
Open Live Terminal
Twin 04
Nemotron-Mini-4B
NVIDIA Nemotron local twin with 32 layers, 32 heads, hidden dim 2048, and GQA Swift compute mode.
Tok2048GQAHead
Open Live Terminal
Claude Revelation
Claude is a Python Orchestrator. So is Everything Else.
Column 01
Claude / Python
Language becomes orchestration: read state, choose action, call tools, seal the result, continue the chain.
prompttoolstateseal
Column 02
Browser ROM
The page is not a brochure. It is the visible control surface that dispatches local runtime work and writes WORM truth.
HTMLJSWORMexport
Column 03
Everything Else
Ollama, Swift, Ada contracts, and shadow builders are all orchestration lanes around one loop: act, prove, seal.
LLMproofAdastone
Runtime Boundary
GitHub Pages serves the arena ROM. Each twin page now includes an Ollama In A Box panel that calls your local Ollama runtime from the browser at http://127.0.0.1:11434 and seals model results into JS WORM. The Python orchestrator remains available for local/manual inspection.