Supermix v88 (v88_corpus.selected.pt)

Standalone milestone release for v88_corpus.selected.pt from kai9987kai/Supermix β€” the 30.2M-parameter (5.69M active/token) MiMoMix MoE + Recursive Thinking Core model bridging supermix-v87 and supermix-v89.

Verified Training & Generalisation Metrics (Windows CPU)

  • Parameters: 30,211,709 total / 5,693,933 active per token (64 routed experts, top-2, 2 MTP speculative draft layers).
  • Selected Checkpoint: step 21,000 (selected_on = accuracy, in-training probe accuracy 0.95, best_dev_loss = 0.025629).
  • Generalisation Tiers: tier1_seen_response loss 0.014598, tier3_unseen_sentence loss 0.195604, all 5/5 v58 generalisation checks passed = True.
  • Receipts included: results/generalisation_results.json, results/supervised_run.json.

Quick Start

pip install torch numpy flask huggingface_hub
huggingface-cli download Kai9987kai/supermix-v88 --local-dir .
python example_usage.py
Downloads last month
-
Inference Providers NEW
This model isn't deployed by any Inference Provider. πŸ™‹ Ask for provider support

Model tree for Kai9987kai/supermix-v88

Finetuned
(1)
this model