Agent-native models, bound by evidence.

Ainfera builds Neptune — models native to the agent loop, produced by a factory that ships only when identity, data rights, pinned harnesses, and sealed evaluation bind.

We build models that finish work in the world —
and only ship them when the evidence holds.

The unit is the loop, not the monologue.
Proof before claim.
Open Core verifies. Closed specialists deliver.
Refusal is a feature.
Class, not cosmos.

Neptune is trained and measured inside agent protocols:

plan → act through tools → read the world → recover → complete

under a pinned harness — not a chat model wrapped in a prompt.

We are a model factory, not a marketplace.
Serving rails distribute; they do not define us.

27B and 9B are independent evidence chains. Neither inherits the other’s scores, quants, or release authority.

public-data admit
  → base reproduction
  → frozen peers + harness pin
  → trajectories + pilots
  → earned SFT → DPO → GRPO (only with uplift)
  → quant + device certificates
  → sealed eval + independent reproduction
  → claim envelope only as wide as the receipts

If the chain is incomplete, the factory holds.
A factory that cannot refuse is a cost sink.

Training completion is not a release criterion.

No Neptune model is announced as superior, public, or fundable on vibes. A release path requires:

1. Preregistered, held-out comparative evidence under a frozen agent harness
2. Provenance-bound reproducibility
3. A named device / serving envelope when deployment is claimed
4. A passing release gate

Claims stay inside a preregistered local-agent class (size, hardware, VRAM, latency, harness, quant) — never unqualified “best model.”

Open Core is verification and distribution.
Commercial depth sits in closed agentic specialists.

• Website: ainfera.ai (in progress)
• Hugging Face: huggingface.co/ainfera-ai
• GitHub org: this profile
• Models: published when certificates clear

Partnerships and questions: via ainfera.ai when live.