Stay
Curious.

chapter 42 once upon a time, a very curious human clicked “Start here”… ✦ ✧ ✦

and found ideas! ✧ ✦ ✧

and found wonder ✧ ✦ ✧

chapter 42 once upon a time, a very curious human clicked “Start here”… ✦ ✧ ✦

Based AI

Labs21 Software. I'm constantly updating the model in hopes of competing with giants along with zero censorship.

RTX 5080 High-Speed Architecture

By leveraging the RTX 5080's massive VRAM bandwidth and specialized Tensor Cores, we achieve unprecedented throughput for large-scale models. This generational leap enables high-parameter execution on consumer workstations with server-class performance.

3-Phase Strategic Evolution

Intelligence isn't built in a day. Our structured methodology evolves models through distinct phases—from basic reasoning and logic foundations to deep, specialized domain knowledge and recursive self-improvement cycles.

Scaling Sustainably

Democratizing AI training means making it affordable. We focus on high-efficiency scaling that respects both your hardware budget and our planet's energy. Intelligence should be accessible without environmental compromise.

New Core

BDH Neural Architecture

Saga is moving to a from-scratch Bounded Dynamic Highway core: a looped neural network that replaces attention, RoPE, and large KV caches with causal bounded hidden dynamics. Instead of keeping every past token, each BDH block keeps a tiny recurrent state and routes signal through gated neural highways.

Bounded state

Kernel size 5 means only the last 4 hidden states are cached per layer.

Grouped branches

12 temporal branches mix nearby history without attention heads.

Looped depth

28 physical blocks run twice for 56 effective neural passes.

Local training

FP8 block math and regional compile keep the run inside RTX 5080 VRAM.

How BDH reads

Small memory, repeated thought.

Related neurons share a tiny recurrent state, then loop through gated paths for another pass.

4 states cached 28 x 2 passes RTX 5080 local