Matrix54
Qwen2.5-3B LoRA CPT, three optimizer families, three LR regimes, three arms, two seeds, and 5,000 updates per run.
Protocol cardPublic technical diligence
Evidence Room
Protocols, machine-readable outcomes, integrity hashes, and limitations behind Qluon's current Optimizer+ claims.
Evidence hierarchy
One primary study, two supporting gates.
Matrix54 is the primary two-seed breadth study. The completed RC1.06 three-pass matrix adds a longer, one-seed comparison across AdamW, Muon, and Lion. Matrix12-v3 is supporting full-parameter transfer evidence. Their protocols and evidentiary strength differ, so the results are kept separate.
RC1.06 three-pass matrix
18/18 runs: AdamW, Muon, and Lion; healthy and degraded learning rates; three arms; one fresh seed; 23,450 updates per run.
Protocol and endpointsMatrix12-v3
Qwen2.5-1.5B full-parameter CPT, three optimizers, healthy and degraded regimes, one development seed.
Protocol cardClaim register
What the completed studies support.
| Claim | Measured result | Comparator | Boundary |
|---|---|---|---|
| Healthy quality preservation | Matrix54 healthy PPL differences from -0.13% to +0.16% | Matched plain public native | 3B LoRA CPT, two seeds |
| Lower training compute | 15.23% to 41.87% lower GPU-hours by optimizer-family mean | Corresponding tested public accelerated runtime | Matrix54 only |
| Lower memory | 12.32% to 14.29% lower peak VRAM by optimizer-family mean | Matched plain public native | Matrix54 only |
| Degraded-run protection | All three degraded regimes improved; Lion showed a large collapse rescue | Matched plain public native at the same LR | Large Lion effect is not a typical improvement estimate |
| Longer-horizon cross-optimizer | Healthy PPL within 0.03%, with 18.01% to 64.15% fewer GPU-hours; degraded PPL 36.66% to 49.01% lower, with 19.83% to 67.02% fewer GPU-hours | Matched tested accelerated native for each optimizer | One seed, 1.5B LoRA CPT; peak VRAM 3.08% to 5.25% higher |
| Muon endurance highlight | Healthy: 4.0543 to 1.4533 GPU-hours (-64.15%); degraded: 4.1352 to 1.3636 GPU-hours (-67.02%) | PyTorch Muon with fused AdamW fallback and compiled training loss | One seed; +5.25% peak VRAM; not a comparison against every optimized Muon stack |
| Full-parameter transfer | 6/6 composite pair gates passed; mean GPU-hours -3.45%, peak VRAM +1.93% | Matched public native arm | One development seed; supporting evidence |
Protocol integrity
Designed for matched comparison.
- 01Matched training state. Comparisons freeze model initialization, data order, update budget, hardware, and evaluation protocol within each pair.
- 02Declared comparators. Quality is compared with matched public native; GPU-hour claims name either public native or the tested accelerated runtime.
- 03Fixed endpoints. Matrix54 and Matrix12-v3 use 5,000 updates and 5.12M input tokens per run. The RC1.06 endurance matrix uses 23,450 updates and 24.01M tokens per run.
- 04Machine-readable evidence. Public JSON and SHA-256 manifests allow a reviewer to check the disclosed records without exposing product internals.
Limitations
Claims stop where the evidence stops.
Universal speedup
Results do not establish the same gain for every model, GPU, framework, optimizer configuration, or distributed topology.
Population-level generality
Matrix54 uses two seeds; the supporting studies use one seed. Independent customer replication is required.
Product internals
Controller rules, actuator coefficients, runtime implementation, weights, and raw data are intentionally excluded.
Large Lion rescue caveat
The -80.11% degraded Lion mean reflects rescue of two severely degraded high-LR native runs. It must not be interpreted as an expected improvement for ordinary Lion training.
External validation
The next gate is a customer-controlled replication.
Matrix54 complete
18 cross-optimizer runs complete
Customer workload and baseline
Pending deployment