03 / Bench
Serious upside. Honest status.
Independent research. Years from finished. We publish signal. We do not invent benchmarks. No production SOTA.
Constraint
Years from finished. BulmaX context is graph-backed, sealed to 5,242,880 tokens with local attention 1024 — not dense 5M context. Beerus is swarm head + Frontier graph. Public artifacts live under Hugging Face / YNSScarSaiyan. A public note covers BulmaX, Beerus, and Goku as architecture directions. Genesis is a separate chip destination — not a model, not E-Cad, not taped out, never CUDA. Research statements — not production SOTA.
Architecture research directions
BulmaX
Graph-backed context sealed to 5,242,880 tokens, with local attention 1024. That is not dense 5M context — do not read it that way. Self-modifying transformers and dynamic neurogenesis are architecture research directions under study, not present shipping capability. Independent research. Years from finished. No production SOTA.
bulmax-gen2-weights ↗Swarm head + Frontier graph
Beerus
Non-transformer stigmergic micro-experts. Swarm head plus Frontier graph. Architecture research, not a product. Years from finished. No SOTA claim.
beerus-b200-checkpoints ↗Long-horizon research
Coherence AI
How a system stays coherent across time — not just the next token. Related long-horizon work lives here. Early. Years from finished. No production claim.
YNSScarSaiyan on Hugging Face ↗Chip destination — not a model
Genesis
Custom AI compute chip destination. Today: twin + formally checked SystemVerilog / RTL, COTS Artix-7 FPGA path. Not taped out. Never CUDA — the chip destination is non-CUDA. Genesis is not a model. Genesis is not E-Cad. Research and silicon path honesty only — no tape-out claim.
Ask about Genesis ↗
Related public work
Goku is part of the same public architecture set. It sits with BulmaX and Beerus as research, not a shipped product. Checkpoints and datasets update on Hugging Face when they are ready to show.
Ready to book