We are announcing the general availability of Aether 1.0 — the first version of our simulation-native foundation model. Aether is trained primarily on simulation trajectories, instrument data and design history rather than on scraped natural-language text. The release includes three sizes (7B dense, 40B dense, and 280B sparse mixture-of-experts with 40B active parameters per token), a 1M-token context window, streaming time-step rollouts, and full capability gating in the model layer.
What this means in practice: Aether is now the substrate behind every Apex discipline — engineering, semiconductors, drug discovery, autonomous labs, and software. The same foundation model handles all five. Customers running pilots that started during the preview programme are migrating today; new pilots start on the GA release.
The 1.0 release is the culmination of roughly eighteen months of training, evaluation and customer-pilot work. The corpus combines petabytes of solver runs across computational fluid dynamics, finite-element analysis, molecular dynamics, density-functional theory and digital RTL simulation, with instrument traces, design history and curated engineering text adding the contextual grounding. Failure archives are upweighted during training; the model learns from disasters as well as from successes.
We are simultaneously publishing the technical paper at /research/simos-1-0. The paper covers the architecture, the training corpus, the per-discipline benchmark results, the cross-domain transfer findings, and the failure modes we ship the model with. The model card, the reproduction kits for every published benchmark, and the calibration tooling are all open under the Research tier.
Headline benchmark numbers: 3.6% mean Nusselt error on the standard impinging-jet benchmark (versus 7.8% for a matched commercial RANS baseline); ≤0.4% worst-negative-slack gap on the public open-cell-library timing benchmark; 1.06 kcal/mol RMSE on the public eight-target FEP benchmark; 0.6% lift and 1.0% drag polar deviation on the standard transonic wing-body benchmark. The full numbers and reproduction kits are linked from the paper.
Pricing for Aether 1.0 follows the tier structure published at /pricing. The Free tier exposes the 7B endpoint with monthly call limits and the public benchmark suite. Plus ($20/month) unlocks the 40B endpoint and every discipline. Pro ($200/month) adds the 280B sparse-MoE configuration, the eval harness, and SDK access. Team ($30/seat/month annual) is the working-team tier with SSO, audit logs and shared workspaces. Enterprise (custom) is for VPC, on-prem and sovereign deployments.
Availability rolls out in stages. Team and Enterprise customers have access today. Pro customers see the 280B sparse-MoE configuration go live this week. Plus and Free tier upgrades complete over the next two weeks; existing customers will receive an email notification when their tier upgrades. Sovereign installs require an Apex forward-deployed engineer to coordinate the upgrade; affected customers have been notified directly.
We are committed to the v1.0 surface area through at least the v1.5 release. Customers depending on specific model behaviour for production deployment can pin the v1.0 model version and the corresponding refusal-corpus version for the duration of their MSA. The upgrade path to 1.1 and 1.5 will be documented six months in advance of each release; deprecation windows are six months for any v1 API breaking change.
The safety architecture is documented in detail at /research/biosecurity-boundaries and /research/refusal-corpus-methodology. The refusal corpus is hash-versioned and pinned to the 1.0 release; customer extensions via the zero-knowledge protocol are supported on the Enterprise tier. We commission external red-team passes quarterly; the first pass against the GA build completed three weeks ago and the findings have been incorporated into the released model.
For customers running pilots: the v1.0 release does not change the pilot programme. Pilots remain three to eight weeks, scoped to one workload, with a written win condition agreed in advance and a forward-deployed Apex engineer paired with the customer team. The pilot fee is creditable against year-one platform cost for customers who proceed.
We will hold a public Q&A on the technical paper this Thursday. Details will be posted at /news/simos-1-0-ga-qa closer to the time. Customers and academic researchers are welcome.
The team that shipped Aether 1.0 is small relative to the scope of the work, and we are hiring across research, platform, applied engineering, design, safety and go-to-market. Open roles are at /careers; we read every inbound and reply within one week.