Supermix v102 (v102_scf_frontier.pt β€” Frontier Architecture & Muon Synthesis)

Standalone release for v102_scf_frontier.pt and the v102 Frontier Evolution runtime from kai9987kai/Supermix (training release ff35c75c).

What v102 Adds

  1. Muon-Trained Frontier Neural Checkpoint (weights/v102_scf_frontier.pt):
    • Warm-started from Kai9987kai/supermix-v101 (v101_scf_cognitive.pt, 9,460-token vocab; 36,628,023 total / 9,161,127 active params) using MuonHybridOptimizer (--optimizer muon_hybrid, 5-step quintic Newton-Schulz orthogonalized Nesterov momentum for 2D hidden matrices + decoupled AdamW for embeddings/norms/biases/gates).
    • Enables Manifold-Constrained Hyper-Connections (use_mhc_residual=True, mhc_streams=2, mhc_sinkhorn_iters=20) projecting residual stream mixing onto the Birkhoff polytope (doubly stochastic, spectral norm <= 1).
    • Enables Active Apical Dendritic Coincidence Gating (cns_dendritic_apical=True) on the bilateral male-CNS ConnectomeCore and Continuous Thought Machine Pairwise Temporal Synchrony (thinking_ctm_synchrony=True) on RecursiveThinkingCore.
    • Achieves best unseen-sentence composition loss in the v101–v103 ladder: tier3_unseen_sentence = 0.0502 (ppl 1.0515), tier1_seen_response = 0.0048 (ppl 1.0048), tier3/tier1 = 1.046x, dev_loss = 0.0181, and all 5/5 v58 generalisation checks [PASS] including exact speculative_matches_greedy.
  2. Six Cross-Ecosystem Frontier Diagnostic Engines (source/nexus_v102_frontier.py, all answer_authority=False with SHA-256 receipts):
    • NexusSearchHybridRRF, TesseraRatifiedLemmaCommons, CounterformFalsificationAuditor (CVaR + fail-closed ABSTAIN_UNSUPPORTED), CarrierInteractionAtlas (exact 2^k Mobius decomposition + GhostInTheMachine v7 R^2 + Supermix-Aether FLOP/KV profiler), SonapseDiscoveryMonitor, and HyperionComputeScheduler.
  3. Expanded Zero-eval Exact Rational Solvers (source/nexus_solver.py):
    • 3x3 Gauss-Jordan systems, 2x2/3x3 determinants & trace, modular exponentiation/inverse, prime factorization, Euler's totient, binomial PMF, Bayes' theorem, parallel resistors, thin-lens optics, and python -m scf frontier --json.

Honest Measured Benchmarks (Windows CPU, max_new_tokens=96)

Evaluation Suite Result Receipt
Generalisation Ladder (v58) dev_loss = 0.0181 (tier1 = 0.0048, tier3 = 0.0502, 5/5 PASS) results/generalisation_results.json
Standalone Neural seed=82 (41 tasks) 35 / 41 (0.8537) results/eval_v102_seed82.json
Standalone Neural seed=83 (82 tasks) 72 / 82 (0.8780) results/eval_v102_n82_seed83.json
Standalone Neural seed=84 (82 tasks) 71 / 82 (0.8659) results/eval_v102_n82_seed84.json
3-Seed Combined Standalone (205 tasks) 178 / 205 (86.8%) Paired against v96 (180/205), v100 (179/205), v101 (178/205)
SCF Hybrid (python -m scf bench --seed 82 --neural) 41 / 41 (1.0000) results/scf_bench_v102_seed82.json

Per Rule 9 promotion gating (automatic_promotion_enabled=False), v96_physio_neurogenesis.selected.pt (Kai9987kai/supermix-v96) remains #1 Promoted Default; v102_scf_frontier.pt is released as Experimental.

Quick Start

pip install torch numpy flask huggingface_hub
hf download Kai9987kai/supermix-v102 --local-dir .
python example_usage.py
python source/scf frontier --json
python source/supermix_chat_server.py --model v102_scf_frontier=weights/v102_scf_frontier.pt --port 8780

Checkpoint discovery runtime and release integration: Supermix 7feb8e0e.

Downloads last month
27
Inference Providers NEW
This model isn't deployed by any Inference Provider. πŸ™‹ Ask for provider support

Model tree for Kai9987kai/supermix-v102