Global Intent Company · Custom Foundation Architecture Research

INTENT RLSE-v2 (187.3M Custom Recurrent Latent Search Engine)

Choose Facts-500 or SFT-1000 below. Facts-500 adds 500 factual updates after SFT-v2 step 750; SFT-1000 is the earlier chat checkpoint. Both are experimental and can give incorrect answers. Features 10 Context Encoder Blocks, a 4-Block Shared-Weight Recurrent Controller unrolled K=4 times with B=3 Parallel Candidate Branches, a 2-Block JEPA World Model, and a Self-Supervised Latent Critic.

Params: 187.3M Recurrent passes: 4 Two checkpoints available
INTENT model checkpoint
Example Research Prompts
RLSE-v2 Live Latent Telemetry
10 Encoder Layers + 4x Recurrent Core (B=3 Branches) + JEPA World Model
0.0 tok/s · 0 toks
Mean Critic Proxy Score
0.000 (0%)
Effective Layer Traversals
10.0 (expected passes: 0.00/4)
Active Control Operator
—
PonderNet / ACT Recurrent Halting Distribution & Winning Candidate Branches
Pass k=1 (Winner: Branch B—) 0% halt weight
Pass k=2 (Winner: Branch B—) 0% halt weight
Pass k=3 (Winner: Branch B—) 0% halt weight
Pass k=4 (Winner: Branch B—) 0% halt weight
Per-Token Critic Score
Awaiting generation...
Inference Profile
0 10
1 100
0.1 1
16 256
1 1.5
0 1
0 6