Text reasoning, structured outputs, and tool use for controlled workflows.
Models built to become
your operating asset.
We develop the model layer ourselves: weights, post-training recipes, evaluation suites, release gates, and versioned model cards. That gives your team a product you can improve—not simply rent.
Open model registry ↓Owned model lifecycle
We ship the
artifact and the method.
Each model has a controlled lineage from data curation to post-training and production evaluation. Customers can select a SPARK release, extend it with their permitted data, and run it inside their chosen boundary.
Document, image, and visual inspection with task-specific evaluation.
Enterprise retrieval, evidence selection, and access-aware knowledge workflows.
MODEL CARDSPublished
MODEL WEIGHTSPrivate evaluation only
TRAINING DATANot publicly released
BENCHMARKSProtocol published; verified results pending
GITHUB / HUGGING FACENo official public repository announced
Improve models
with a closed loop.
One control plane,
more model choices.
SPARK models remain the core of the platform. When a workflow benefits from an approved third-party model, connect it through official access and operate it with the same guardrails.
Authorize
Connect an approved provider account or approved model artifact under its official terms.
Evaluate
Run the model through the same quality, safety, latency, and cost evaluation gates.
Route
Add it to policy-based routing beside SPARK models and your own adapted releases.
Operate
Observe its usage, trace decisions, and apply your organization’s access controls.
Provider names, trademarks, models, availability, and commercial terms are only shown after confirmed authorization. The platform capability is real; specific integrations require the appropriate provider agreement.