Models you own.
Open-weight domain models from the Neural Infrastructure family, published on Hugging Face, proven on our own live deployments, and yours to run anywhere: our fleet, your cloud, or your own hardware.
Rented intelligence is a dependency. Owned intelligence is an asset.
API models are someone else’s weights, someone else’s terms, someone else’s price changes. An owned domain model is a file you hold: auditable, portable, and running wherever your data lives. It’s layer 05 of the stack, made literal.
Weights you hold
Apache-2.0 licensed adapters you can download, inspect, and serve on any infrastructure.
Provenance you can audit
Every release ships with its evaluation results, and every training run is leak-tested so no client or brand data rides along.
Run at your altitude
Serve them on our managed fleet, inside your VPC on the Own tier, or on your own keys and hardware.
Domain brains, not general chatbots.
Each model is trained for one job an AI employee actually does, and evaluated blind against its own base model before it ships.
B2B Marketing v2
The copywriting brain behind our marketing AI employees. Positioning briefs in, on-brand B2B assets out.
- LoRA adapter on Qwen3-32B · Apache-2.0
- Blind-judged 7.16 vs 4.79 for its base model
- Preferred head-to-head over base in 86% of matchups
- Fabricates fewer statistics than its own base
- Zero-leak verified, no client or brand data inside
B2B Marketing (7B)
The compact first release. The same B2B copy discipline, light enough for a single modest GPU.
- LoRA adapter on Qwen2.5-7B-Instruct · Apache-2.0
- 13,000+ curated, brand-scrubbed training pairs
- Zero-leak verified across generation tests
- Suited to on-device and single-GPU serving
CFO Brain v1
An advisory finance brain for the Finance Manager AI employee: judgment frameworks for Indian mid-market companies, not your books.
- LoRA adapter on Qwen3-32B · Apache-2.0
- Blind-judged 5.50 vs 3.79 for its base model
- Preferred head-to-head over base in 75% of scenarios
- Grounded arithmetic, calibrated refusal on current rates
- 1,926 verified training pairs · zero-leak verified
This model wrote this website.
The marketing model’s acceptance test was real work: drafting new landing copy for this site from our positioning brief. The drafts were judged a significant improvement on the copy they replaced, refined by a human editor, and shipped. The hero you read on the way in came out of that session.
Evaluation sets are published too: b2b-marketing-eval and cfo-advisory-eval. How they are built is on the Datasets page.
The dataset is the moat. The method is the guarantee.
Base models are a commodity. The value is in the training data and the discipline around it.
Curated pairs, not scraped sludge
Training data is generated against real B2B scenarios, then quality-filtered by calibrated judges. The weak majority gets cut before training starts.
Blind evaluation before release
Every model is scored blind against its own base across held-out prompts. If it doesn’t beat the base, it doesn’t ship.
Anti-fabrication testing
Outputs are audited for invented statistics and unsupported claims, a failure mode we test for explicitly, not hope away.
Scrubbed and leak-tested, every run
Client and brand data is scrubbed before training and verified absent after it. Public weights carry zero customer material.
Put an owned model to work.
We’ll map which of your workflows deserve an owned domain model, and how to serve it at your altitude. No pitch deck. Just a technical conversation.