HELIX-200M-F35B

A 209M-parameter H-E-R hybrid can be pretrained on a 40B-token Chinchilla-matched corpus entirely on one M3 Max, from scratch after the 127M generation was ruled parameter-bound.

HER×5 (15 layers), d_model 1024, rotation-HSL (r_max 256), MoE top-2 of 4. Corpus: corpus_200m_diverse — FineWeb-Edu 50% / DCLM 18% / math 8% / Wikipedia 7% / books 5%, 40,000,079,759 tokens (uint16). LR 3e-4, 2,000 warmup, 16,384 tok/step, BF16.

• ◆Scale (209M vs prior 57M/127M)
• ◆Corpus size and mixture

• →Step 120,140: train loss 2.824 (PPL 16.85), validation PPL 39.4 — plateau band 39.1–40.3, best ~39.09 @ step 118,000.
• →First benchmark at step 44,673: Open-SLM-style avg 40.4%.
• →Throughput ~2,900 tok/s sustained; grouped-GEMM MoE adopted mid-run for +13.7% tok/s.

• ·The 2026-07-06 decision memo called the remaining 97.5% of the 40B-token autopilot schedule the program's 'single biggest waste risk' given the plateau.
• ·Step 120,000 was deliberately protected as a checkpoint and the trunk run was branched, not abandoned as a failure.

Deliberately branched at step 120,140 rather than continued — a scheduling decision, not a failed run. The plateau turned out to be caused by the schedule, not the architecture (see HELIX-200M-C1X-RECOVERY).

Branch-and-decay: resume from the protected checkpoint with a real WSD schedule.