Healthy condition
Quality nearly unchanged
Qluon / AI Training Control
Reduce GPU Cost.
Protect Model Quality.
Optimizer+ adds intelligent control around the optimizer you already use.
3 Founding Design Partner SlotsBring one workload. Native vs Qluon.
Training stateObserve model, data, system
Qluon controlAnalyze · Decide · Regulate
Your optimizerAdamW · Muon · LionMeasured results / Muon three-pass study
About one-third the GPU-hours.
On the matched 1.5B LoRA CPT workload, Optimizer+ reduced total GPU-hours in both Muon regimes. Healthy validation PPL was nearly unchanged; degraded validation PPL improved.
Review the complete study-64.15%GPU-hours
Degraded condition
Quality improved
Shared scale: 0 to 4.2 GPU-hours. Lower is better.
Scope: Qwen2.5-1.5B, LoRA CPT. 23,450 updates; seed 4007. Comparator: the tested accelerated native Muon implementation, not every optimized Muon stack. Peak VRAM was 5.25% higher with Optimizer+. One internal seed is a pilot signal, not a general guarantee.
Breadth behind the result
Evidence at scale,
not a single run.
Matrix54 adds matched evidence across optimizer families.
- Completed Matrix54 runs
- 54
- Optimizer families
- 3
- Seeds per condition
- 2
Matrix54
PrimaryCross-optimizer breadth
- Workload
- Qwen2.5-3B · LoRA CPT
- Design
- 54 runs · 3 optimizers · 2 seeds
- Budget
- 5,000 updates per run
- Boundary
- Internal, single GPU
RC1.06 endurance
SupportingLong-horizon matched evidence
- Workload
- Qwen2.5-1.5B · LoRA CPT
- Design
- 18 runs · 3 optimizers · 1 seed
- Budget
- 23,450 updates per run
- Boundary
- One model, workload, and seed
Matrix12-v3
SupportingFull-parameter transfer gate
- Workload
- Qwen2.5-1.5B · Full CPT
- Design
- 12 runs · 6 pairs · 1 seed
- Budget
- 5,000 updates per run
- Boundary
- Balanced-memory development gate
AdamW: healthy and degraded resultsA second optimizer from the same endurance study
AdamW endurance result
Less compute in both conditions. Less quality loss under stress.
This AdamW comparison is now part of a completed 18-run endurance study across AdamW, Muon, and Lion.
Inspect all six comparisons| Regime | Validation PPL native → Qluon | PPL change | GPU-hours change |
|---|---|---|---|
| Healthy | 12.4788 → 12.4824 | +0.03% | -18.01% |
| Degraded | 22.3476 → 14.1541 | -36.66% | -19.83% |
Qwen2.5-1.5B LoRA CPT, seed 4007. Comparator: matched public accelerated AdamW. Lower PPL and GPU-hours are better; one seed is a pilot signal, not a general guarantee.
Technology
Optimizer design is only half the system.
Qluon adds a control layer that observes, analyzes, and regulates training in real time.
Traditional training selects optimizer behavior before the run. Optimizer+ adds a second dimension: state-dependent control during the run.
It wraps the optimizer you already use. The base optimizer remains recognizable; Qluon decides when to preserve its update and when to apply bounded regulation.
01Your optimizerAdamW, Muon, or Lion
02 / Qluon controlOptimizer+Observe, decide, regulateSensor · Analyzer · Controller · Actuator
03Training stepLogged and reversible
Next: evidence, security & compatibility Explore the control architecture
Integration
Simple to integrate.
One wrapper. Your loop stays familiar. Wrap your existing optimizer with Optimizer+ in a few lines of code.
The pilot package provides a controlled integration path, runtime fallbacks, deterministic checkpoints, and customer-owned telemetry.
Request the integration guideimport torch
from learnbywire import OptimizerPlus
base = torch.optim.AdamW(model.parameters(), lr=lr)
optimizer = OptimizerPlus(base, model=model)
for batch in train_loader:
optimizer.zero_grad()
loss = model(**batch).loss
loss.backward()
optimizer.step()Financial view
Translate GPU-hours into an accountable pilot target.
The model is deliberately conservative and transparent. Adjust only the share of annual training spend in scope and the saving you want the pilot to prove.
Estimated saving = annual GPU spend × eligible workload × measured reduction
Illustrative planning scenario, not a forecast or guarantee. Savings depend on hardware, workload, baseline maturity, utilization, and deployment scope.
$
Illustrative annual saving$150,000Equivalent to $12,500 per month
Build our measured business case
Next: test your workload against Qluon
Operations view
Two weeks to a measured answer.
Run inside your environment, against your baseline, with predeclared gates. Your result determines the next step.
Pilot scope, process, and workload fitReview the four stages and generate a qualification brief
Freeze the baseline
Agree workload, native comparator, quality target, budget, hardware, and success gates.
Integrate privately
Wrap the optimizer, validate fallback and checkpoint behavior, and keep data in your environment.
Run matched evidence
Compare quality, GPU-hours to matched quality, stability, and operational overhead.
Make the decision
Receive an evidence pack: pass, conditional pass, or stop. No vague success criteria.
Pilot fit
Map your workload to the current evidence.
This check runs locally in the browser and sends no workload information anywhere.
Direct evidence matchStart with a matched two-week qualification.
Matrix54 directly covers 3B LoRA CPT with AdamW, Muon, and Lion. The pilot still uses your accepted baseline and success gate.
Reduce training cost without replacing your optimizer.
Discuss a customer pilotAdd a differentiated control layer to your training stack.
Explore a partnershipEvaluate the evidence, integration boundary, and enterprise route.
Meet the founderAbout Qluon
Building the control layer for AI training.
Qluon is an AI infrastructure company with a presence in Delaware and Ghent, Belgium. Our focus is how models, optimizers, and compute work together during training.
A strong starting configuration is only part of the job. Training behavior changes as a run progresses. Qluon is building the systems that help teams observe those changes, respond, and measure the effect on model quality and compute cost.
LearnByWire is our training-control architecture and platform direction. Optimizer+ is the first pilot product, combining optimizer control and runtime execution around the tools a team already uses.
Talk with the founder- Founder
- Anis Radianis
View LinkedIn profile - Locations
- Delaware, United States
Ghent, Belgium - How we work
- Private customer pilots with an agreed baseline, matched evaluation, and a clear decision. Training runs inside the customer's environment.
- Current stage
- Internal cross-optimizer evidence is complete. Independent customer replication is the next milestone.
- Intellectual property
- Patent pending. A provisional application has been filed.
Go deeper
Technical and company detail.
Full benchmark resultsMatrix54, endurance, trajectories, and full-parameter transfer
Research coverage
Tested beyond one training recipe.
Qluon's development program spans multiple training objectives, adaptation methods, model scales, optimizer families, and operating regimes.
- Training
- PT, CPT, SFT
- Adaptation
- LoRA, QLoRA, full parameter
- Scale
- 1.5B and 3B models
- Optimizers
- AdamW, Muon, Lion
- Regimes
- Healthy, stressed, degraded
Coverage means evaluated in the research program. Evidence strength varies by cell; headline claims remain tied to the named frozen protocol beside each result.
Measured evidence
Inspect the evidence, not just the headline.
Matrix54 is the primary breadth study. Explore its matched two-seed results by optimizer and operating regime; the comparator is stated beside every metric.
Matrix54 remains the primary two-seed breadth study. The separate RC1.06 endurance evidence covers AdamW, Muon, and Lion at one fresh seed over 23,450 updates; Matrix12-v3 is a separate full-parameter transfer study.
Evidence passportMatrix54
- Workload
- Qwen2.5-3B LoRA CPT
- Design
- 3 optimizers × 3 regimes × 3 arms × 2 seeds
- Budget
- 5,000 updates / 5.12M input tokens per run
- Comparators
- Plain public native and tested public accelerated runtime
- Integrity
- Matched initialization, data order, steps, and endpoint rule
- Boundary
- Internal single-GPU evidence; not a universal performance guarantee
Evidence passportMatrix12-v3
- Workload
- Qwen2.5-1.5B full-parameter CPT
- Design
- 3 optimizers × healthy/degraded × 2 arms
- Budget
- 5,000 updates / 5.12M input tokens per run
- Result
- 6/6 composite pair gates passed
- Runtime profile
- Balanced memory; mean peak-VRAM change +1.93%
- Boundary
- One development seed; supports pilot replication, not a population claim
Evidence passportRC1.06 three-pass endurance
- Workload
- Qwen2.5-1.5B LoRA CPT
- Design
- 3 optimizers × healthy/degraded × eager/accelerated/Optimizer+; 18/18 complete
- Budget
- 23,450 updates / 24.01M input tokens / 3.0016 passes
- Seed
- Fresh preregistered seed 4007
- Primary comparator
- Matched tested accelerated native for each optimizer
- Boundary
- One seed, one model and workload; peak VRAM rose 3.08% to 5.25% versus accelerated native
54completed runs
18matched three-arm blocks
2seeds per optimizer/regime
3optimizer families
- Model
- Qwen2.5-3B LoRA
- LR
- 5e-5
- Budget
- 5,000 updates / 5.12M tokens
- Seeds
- 2203 and 2213
Reading rule: negative is lower/better. PPL and VRAM compare with matched plain native; GPU-hours compare with the optimizer's tested accelerated public runtime. Percentages are equal-seed descriptive means.
Cross-optimizer evidence
One control concept, three optimizer families.
Matrix54 family/regime means. PPL compares with matched plain native; GPU-hours compare with each tested accelerated public runtime.
| Optimizer | Healthy PPL | Stressed PPL | Degraded PPL | GPU-hours vs accelerated | Peak VRAM vs plain native |
|---|---|---|---|---|---|
| AdamW | -0.13% | -0.66% | -2.68% | -16.97% | -12.32% |
| Muon | -0.04% | -0.48% | -1.31% | -41.87% | -12.87% |
| Lion | +0.16% | -0.33% | -80.11%* | -15.23% | -14.29% |
*The degraded Lion mean is a rescue of two severely degraded high-LR native runs, not a typical Lion improvement. Matrix54 is internal single-GPU evidence on one model/workload and two seeds; it motivates matched customer replication.
Fresh-seed endurance evidence
Three optimizer families. Healthy quality held; degraded quality improved.
The completed AdamW gate and its frozen Muon/Lion extension used the same RC1.06 package, seed 4007, 23,450 updates, and 24.01M input tokens per run. The primary comparison is each family's tested accelerated native arm.
| Optimizer | Regime / LR | Accelerated PPL | Optimizer+ PPL | PPL change | GPU-hours change | Peak VRAM change |
|---|---|---|---|---|---|---|
| AdamW | Healthy / 5e-5 | 12.4788 | 12.4824 | +0.03% | -18.01% | +3.08% |
| AdamW | Degraded / 1e-3 | 22.3476 | 14.1541 | -36.66% | -19.83% | +3.08% |
| Muon | Healthy / 3e-5 | 12.1651 | 12.1684 | +0.03% | -64.15% | +5.25% |
| Muon | Degraded / 1e-3 | 25.0575 | 13.8422 | -44.76% | -67.02% | +5.25% |
| Lion | Healthy / 3e-6 | 11.9749 | 11.9746 | 0.00% | -21.06% | +3.29% |
| Lion | Degraded / 1e-4 | 24.3643 | 12.4244 | -49.01% | -22.98% | +3.29% |
Final larger-sample validation PPL and total process GPU-hours are shown; lower is better. Percentages compare with the matched tested accelerated native arm. Healthy PPL differences were within 0.03%; degraded PPL was 36.66% to 49.01% lower. This is one seed on Qwen2.5-1.5B LoRA CPT, not a universal gain. Download all endpoint metrics and comparator details.
Real trajectory replay
Follow quality and compute through 23,250 updates.
Sanitized online validation telemetry from the completed fresh-seed AdamW endurance study.
The replay uses the fixed online validation sample every 250 updates. The larger frozen endpoint evaluation is reported in the endurance table above. Download sanitized trajectory JSON.
Separate transfer gate
Full-parameter control passed its balanced-memory gate.
Matrix12-v3 tested Qwen2.5-1.5B full-parameter CPT across AdamW, Muon, and Lion. All six matched healthy/degraded pairs passed the preregistered quality, compute, numerical-validity, and peak-memory criteria.
- Runs
- 12/12 complete
- Pair gates
- 6/6 passed
- Mean GPU-hours
- -3.45%
- Mean peak VRAM
- +1.93%
Matrix54 complete
18 cross-optimizer runs complete
Next evidence milestone
Pending customer deployment
SACAR architecture and componentsHow Opt+, Run+, Safety+, and Data+ research fit together
Technical architecture
SACAR: a control loop around training.
Sensor, Analyzer, Controller, Actuator, and Runtime separate observation from action. The result is measurable, bounded, and compatible with the optimizer already in the training stack.
Opt+
Optimizer-state supervision and bounded intervention.
Run+
Compiled execution, supported kernels, caching, and fallback.
Safety+
Guardrails, observability, deterministic checkpoints, rollback.
Data+ research
Evidence-gated work; not included in the current pilot claim.
From training metrics to end-user valueHow pilot results connect to application evaluation
End-user impact
Training control matters only when the model serves people better.
Optimizer+ acts during training, but the final test is downstream: does the selected checkpoint improve the product experience at an acceptable cost?
Every pilot therefore connects training metrics to the customer's own application evaluation before an end-user claim is made.
More dependable quality
Preserve useful learning and reduce the risk that an aggressive run ends in a poor checkpoint.
Faster model iteration
Lower training cost can let product teams test, refresh, and improve models more frequently.
Measured product impact
Connect PPL and retention to domain metrics such as task accuracy, reliability, and user acceptance.
Training state→Optimizer+ control→Validated checkpoint→Application evaluation→End-user release
PPL is a training-quality measure, not an end-user KPI. Qluon does not infer user benefit from PPL alone.
Company strategy and technology landscapePlatform direction, adjacent tools, and the investment case
For Qluon investors
The investment case for a training control layer.
This section is for investors evaluating Qluon as a technology company. Optimizer+ is the initial product wedge: a supervisory layer that enters through the optimizer interface and can expand into a broader AI training control plane.
Unsloth
Optimized kernels, packing, and memory efficiency for supported model workflows.
ComplementaryDeepSpeed / FSDP
Sharding, offload, communication, and distributed memory scaling.
Infrastructure beneath controlVizier / Optuna
Choose hyperparameters across separate experiments using black-box optimization.
Between-run optimizationW&B / MLflow
Track, compare, and govern experiment evidence and artifacts.
Evidence and monitoringQluon Optimizer+
Observe optimizer state, choose a bounded action, and execute a controlled update during training.
The control layerOptimizer+ introduces within-run supervisory control without asking an AI team to replace its model, data, or optimizer.
SACAR combines state observation, bounded actuation, runtime execution, safety, and evidence governance. Exact controller and runtime implementations remain proprietary.
Paid private pilots establish workload-specific value, followed by enterprise deployment and support inside the customer's environment.
Evidence todayCross-optimizer breadth plus endurance
Matrix54, a separate 18-run three-pass endurance study, and full-parameter transfer evidence establish the current technical basis.
Next value milestoneIndependent customer replication
A matched paid pilot tests quality and GPU-hours against the customer's accepted production baseline.
Platform directionLearnByWire Training Control
Request the investor brief
Optimizer+ and Run+ are the initial controlled components; additional layers advance only after their evidence gates pass.
Product names belong to their respective owners. Categories identify adjacent tools, not certified integrations or superiority claims. Category references: Unsloth, DeepSpeed, Vizier, and W&B.
World Summit AI · Amsterdam · 7-8 October 2026
Meet Qluon at booth S16.
Find us in the Startup Accelerator. Bring one training workload and its quality and cost goals; we will review the evidence and define a matched pilot with your accepted baseline.