Skip to main content
MANIFESTO • PHILOSOPHY • ROADMAP

Open Frontier Infrastructure

We believe the frontier of intelligence should not be locked behind closed APIs. Solstice-AI builds open datasets, model weights, and distillation systems that empower researchers and engineers globally.

OPERATING PRINCIPLES

Why We Build in the Open

Brute-force scale is failing. Efficient distillation on verified consensus reasoning is how intelligence democratizes.

01METHODOLOGY

Multi-Teacher Neutrality

Distilling from a single proprietary model inevitably inherits its unique blind spots, stylistic tropes, and safety hallucinations. We curate balanced consensus traces across 8 diverse frontier architectures to train resilient student models.

02OPEN SCIENCE

Radical Artifact Transparency

We do not publish gated previews or selective cherry-picked benchmarks. Every release includes complete Apache Parquet dataset archives, uncensored chain-of-thought scratchpads, training recipes, and AGPL-3.0 model weights.

03SOVEREIGNTY

Sub-8B Local Hardware Sovereignty

True intelligence democratization requires running frontier reasoning on commodity local hardware. We optimize our distillation pipelines specifically for sub-8B parameter architectures that run comfortably on single consumer GPUs.

04VERIFICATION

Step-Level Verification

More data does not equal better reasoning. Every trace in our corpora passes automated compiler execution, mathematical assertion checks, and step-level backtracking validation before inclusion.

EXECUTION TRAJECTORY

Infrastructure Roadmap

2026 — 2027 Post-Training Milestones

AUG 2026RELEASED

Solace 1.0 Omni & Axiom Distillation Corpora

12,586,893 unique conversations (137.7 GB JSONL, 60 source datasets) and Project Axiom (102 GB CoT, 27B+ tokens) across 7 frontier architectures. Exact SHA256 deduplicated, published under AGPL-3.0 on Hugging Face Hub.

Q4 2026IN PROGRESS

Solace-Sub8B Checkpoints (FP8 & AWQ)

Release of calibrated 3.8B and 7B reasoning checkpoints fine-tuned on Solace 1.0 activation datasets. Native inference support for vLLM, llama.cpp, and Apple MLX.

Q1 2027PLANNED

Equinox: Continuous Distillation Harness

Automated pipelines for streaming multi-teacher consensus traces directly into online fine-tuning clusters with continuous compiler assertion feedback.

Q2 2027PLANNED

Formal Mathematical Verification Sandbox

Integration of Lean 4 and Isabelle formal proof checkers for automated step-by-step mathematical reasoning validation in synthetic datasets.

INQUIRIES & COLLABORATION

Collaborate With Solstice-AI

If you are evaluating Solace traces for your research lab, running benchmarks, or interested in collaborating on open distillation infrastructure, reach out directly.