Aether.
The frontier foundation model for the physical world.
Aether is the foundation model Apex trains end-to-end on simulation trajectories, instrument data and design history. It runs forward in time across fluid, structural, electromagnetic, quantum, biological and digital systems — on the same set of weights — and does the work the chat-trained frontier can describe but not simulate: designing molecules, chips, vehicles, processes and campaigns that have to work in the world.
- Modules
- 243
- Disciplines
- 5+
- Context
- 1M
- Status
- v1.0 GA
The every discipline — one model.
The same weights serve every discipline below. Each link goes to the capability page for that area, with the agents, benchmarks and incumbents it replaces.
- Aether · Digital twins
Digital twins
See discipline - Aether · Generative design
Generative design
See discipline - Aether · AppSec
Application security
See discipline - Aether · XDR
XDR / security ops
See discipline - Aether · Container security
Container security
See discipline - Aether · Threat intelligence
Threat intelligence
See discipline - Aether · OT & IoT security
OT & IoT security
See discipline - Aether · Data security
Data security & DLP
See discipline - Aether · Zero Trust / SASE
Zero Trust / SASE
See discipline - Aether · Engineering
Engineering simulation
See discipline - Aether · Semiconductor
Semiconductor design
See discipline - Aether · Drug discovery
Drug discovery
See discipline - Aether · Wet-lab automation
Autonomous wet-lab
See discipline - Aether · Software
Software engineering
See discipline - Aether · Robotics
Robotics
See discipline - Aether · Cybersecurity
Cybersecurity
See discipline - Aether · Cyber-physical
Cyber-physical security
See discipline
The same model, doing the work.
Eight panels of what Aether is already running. Same weights behind every one — ascent CFD, tokamak plasma, drug docking, VQE, hypersonic re-entry, robot planning, battery thermal, materials discovery. The chat-trained frontier can describe these; Aether simulates them forward in time.
- Launch & re-entry
Mach 2.4 ascent · 1.8% drag error vs flight
- Tokamak plasma
MHD equilibrium · 18 MW/m² divertor flux
- Drug discovery
ΔG = −9.2 kcal/mol · FEP-validated
- Materials discovery
DFT · 10⁶ candidates · 36 meV/atom MAE
- Quantum eigensolver
VQE · 64 qubits · 1 mHa chemical accuracy
- Hypersonic re-entry
Mach 8 · 3.2 MW/m² stagnation flux
- Robotics
6-DOF planning · 142 Nm torque · 8 kg payload
- Battery thermal
P2D + pack · 54 °C T_max · 4C fast charge
Ten numbers you can rerun.
A slice of the published benchmark results across the Aether 1.0 model family. Every number ships with a reproduction kit — the case, the seed, the model checkpoint hash. If it stops reproducing, we treat it as a P1 bug.
- 3.6%Cooper impinging-jet · mean Nu error · 40B
- ≤1.0%wing-body cruise drag polar · transonic
- 0.92notched-bar fatigue · Pearson vs measured
- 1.06 kcal/molFEP RMSE · 8-target panel · Aether Bio 12B
- 0.91hERG ROC-AUC · external curated set
- ≤0.4%WNS gap · 7nm predictive PDK · Aether Silicon
- 1.06×perf at iso-area · vs commercial EDA baseline
- 53.6%SWE-bench Verified resolution rate
- R² = 0.79cell tracking · public CTC dataset
- 92%active-learning · conformal coverage · live instruments
Every modality in. Every discipline out.
CAD, mesh, RTL, molecules, spectra, plate data, code and sensor streams flow into one set of weights — and out through discipline heads that emit the artefact each field actually consumes. Same backbone, shared training, one evaluation suite.
The six things Aether does differently.
These are the engineering choices that make one model viable across so many disciplines. None of them are magic.
Unified, not multi-modal
One set of weights spans CFD, FEA, EM, MD, DFT, RTL and digital twin telemetry — coupled by shared latent state, not glued by adapters.
Simulation, instruments, design history
Petabytes of solver runs, sensor traces, lab notebooks and design databases — curated for physical fidelity, not for chat.
Runs where your data lives
On-prem, in your VPC, or air-gapped. No third-party round-trips for IP-sensitive work.
Long-rollout stability
Multi-step trajectories without the drift that plagues autoregressive surrogates. Validated against held-out experimental endpoints.
Solvers as muscle, model as brain
Aether calls the right numerical solver when called-for and amortises it everywhere else. Speed without losing the ground truth.
Built-in biosecurity & export-control
Refuses dangerous syntheses, gates ITAR-class work, and ships with audit trails the legal team actually likes.
What we feed it — and what we don't.
Most of Aether's pretraining corpus is data the rest of the model industry doesn't have access to. The list below is what makes the model what it is.
Petabytes of CFD, FEA, EM, MD, DFT
Computed on internal infrastructure and curated for physical fidelity — not scraped from preprints.
Real wet-lab and bench data
Plate-reader curves, mass-spec runs, imaging stacks, sensor streams from manufacturing lines.
Decades of validated CAD, RTL, schematics
Geometry, layouts and netlists from real, shipped products. Where licences allow, the as-built shape lives alongside the as-designed.
Manuals, standards, papers
ASME, ASTM, IEC, IEEE — read as authoritative grounding, not as a substitute for the simulation data above.
What went wrong, and why
Incident reports, recall databases, postmortems. Failure data is up-weighted; Aether learns from disasters as well as successes.
Permissively-licensed source
For the software-engineering capability. Filtered for licence and quality; we do not train on viral-copyleft code.
Five variants. One corpus.
Every variant shares the training corpus and the eval suite. The choices are about cost and deployment shape, not about which checkpoint is the smart one.
Aether 7B
GADenseParams 7BContext 1MEdge, embedded, sovereign light deploys
Aether 40B
GADenseParams 40BContext 1MMost production workloads
Aether 280B sparse-MoE
GASparse-MoEParams 280B (40B active)Context 1MFrontier accuracy on hardest workloads
Aether Bio 12B
GASpecialisedParams 12BContext 512kWet-lab, ADMET, generative chemistry
Aether Edge 1.3B
PreviewSpecialisedParams 1.3BContext 128kEmbedded sensors, real-time control
Four shapes. One runtime.
The runtime is identical across deployment shapes. The only difference is who owns the network and the hardware.
Apex-managed cloud
SOC 2 Type II environment, multi-region, private networking optional. Best for fast pilots and teams without a strong infra requirement.
Customer VPC
Deploys into your AWS, GCP, Azure or OCI account. Data plane and control plane are yours; the model and the runtime are ours.
Customer data centre
Installs onto your hardware. Supports H100 / H200 / MI300X / Trainium. Quarterly model refresh on your release cadence.
Sovereign, ITAR-clean
Zero outbound traffic. Weights pinned, refusal corpus verified per upgrade. For national-lab and export-controlled work.
The honest specifics.
We publish what Aether is good at, what it is not yet good at, and how we measure both. Reproduce our benchmarks before you bet a roadmap on us.
Dense + MoE configurations
Three sizes: 7B dense, 40B dense, 280B sparse-MoE. Same training corpus, same evals.
1M tokens, including grids
Hold a full transient CFD case, a chip RTL hierarchy, or a year of lab notebooks in context.
Streaming trajectories
Returns time-step-by-time-step — so you can interrupt, branch and steer.
Held-out experiments
We measure against measurements, not against other models. Numbers are in the research log.
Where it is weakest
Highly novel chemistries, rare-event reliability, and anything our training corpus under-represents.
Commercial & research
Commercial deploys are paid; academic and safety-research access is free and tracked.
Answers to the things people ask first.
If your question is not here, write to hello@apexworldlabs.com. A real engineer answers, not a sequence of canned auto-replies.
Aether is pretrained on simulation trajectories, instrument data and design history — not on scraped text. The model learns dynamics rather than prose. Where a chat model approximates 'what an engineer would write,' Aether approximates 'what the system would do.'
No. Customer prompts, code and experiment data are never used to train base models. Customer fine-tunes are scoped to the customer that requested them and are never re-used across tenants.
Cross-domain transfer is one of the strongest empirical results from our training programme. Pretraining on CFD measurably improves performance on cell-painting and ADMET. We think the shared latent is governing dynamics across scales.
First-class. Aether calls the leading open-source solvers across fluid dynamics, structural analysis, electromagnetics, molecular dynamics and chip-design infrastructure as tools. Tool calls are auditable, gated by policy, and observable in the runtime.
Three things. First, ground truth is cheap when a real solver call resolves the disagreement. Second, the failure mode goes into our regression suite. Third, customer-specific fine-tunes can address consistent gaps in the discipline you care about.
Apex-managed cloud, your VPC, on-prem on your hardware, or air-gapped. The runtime is identical; the only difference is who owns the network and the metal.
A foundation model that does the work, not just the chat about it.
Aether is built to think in physics, chemistry, biology and silicon — the substrate of the physical world. We are training the model that designs a rocket today and runs a lab tomorrow. AGI through the disciplines that have to face reality, not through more pages of scraped text.
- AI scientists
Not chatbots. Plans, runs and reports the work.
- One model
The same weights across every discipline we serve.
- Reality is the regulariser
Trained on simulation, instruments, design history — not scraped text.
The work behind this product.
We publish what we learn. The posts below are the substantive notes behind this product — methods, evaluations, case studies.
Start with one hard problem.
Send us the workload that hurts most — the one with a six-figure annual licence, or the one your senior engineers spend their weekends on. We'll show you what Aether does with it.