What is this

TLDR

We published a federated AI orchestration protocol (NSED) that coordinates multiple independent models across physically separated GPUs through structured deliberation rounds with self-evaluation and quadratic voting consensus. No tensor parallelism required.

• Paper: arXiv 2601.16863

• Code: github.com/peeramid-labs/nsed (source-available under BSL 1.1; Agent SDK under MIT)

• Results: 4x lower CapEx matching frontier model quality on AIME’25 and LiveCodeBench. Alternatively framed: 15–30% improved reasoning performance over base models through orchestration alone.

• Protocol economic design (draft): DIP — Decentralised Intellectual Property

We’re looking for research collaborators and early design partners — both to run NSED against domain-specific benchmarks and to stress-test the DIP protocol’s economic design under adversarial conditions.

---

Background: The Physical Layer Problem in Decentralized Coordination

For those who have followed my prior posts here, I’ve been raising a specific concern since 2020: most decentralized protocol design treats the physical telecommunication layer as a solved problem, when it fundamentally isn’t. Software engineers and mathematicians designing these systems rarely account for TCP/IP-level constraints on what “decentralized” can actually mean in practice. I’ve tried to develop this argument across several posts:

• A Local-Node-Favoring Delta to the Scaling Roadmap

• Economics for P2P RPC Data Marketplace

• IPv6 vs Ethereum

• EIP PR #9832

The common thread: if we want credibly decentralized systems, we need to reason about physical distribution of compute, not just logical distribution of consensus. This is the starting point for the work below.

---

NSED: What We Built and What It Proves

NSED (N-Way Self-Evaluating Deliberation) is a Mixture of Models (MoM) architecture where multiple independent AI models from different providers deliberate through structured rounds of proposals, cross-evaluation, and quadratic-voting consensus. The key architectural property: models run on physically separate, federated GPUs with no shared memory and no tensor parallelism. Coordination happens through a structured message-passing protocol over NATS, not through parameter-level synchronization.

Results

| Configuration | Hardware | AIME’25 Score | vs. Frontier | | ------------------------------------------------------------------------------- | ------------------- | ----------------- | ---------------------------- | | NSED Consumer (3 small open-weight models: GPT-OSS-20B, Qwen-3-8B, Gemma-3-12B) | Consumer-grade GPUs | 84% | Matches DeepSeek-R1 | | NSED High-Perf (open-weight models, self-hostable) | Mid-range | 90.0% | Within 0.8% of GPT-5 (90.8%) |

The CapEx comparison: NSED Consumer achieves equivalent reasoning quality at roughly 4x lower hardware cost than running frontier models directly. The mechanism is deliberation diversity — multiple weaker models cross-checking each other produce better results than a single stronger model reasoning alone.

Why This Matters for Ethereum

The dominant AI paradigm (Mixture of Experts) relies on monolithic models with internal routing — fundamentally centralized compute. MoM is a different scaling axis: quality improves through model diversity and deliberation rounds rather than parameter count. This maps naturally onto physically distributed infrastructure where no single operator controls all participants. If you believe Ethereum’s long-term value includes being coordination infrastructure for AI compute, the question is: what does the coordination protocol look like? NSED is one answer — not the only one, but a concrete, benchmarked one.

Licensing Note

NSED’s orchestrator is source-available under BSL 1.1 (free for entities under $1M ARR, academic and research use, and development/testing; converts to AGPL-3.0 after 4 years). The Agent SDK is MIT — zero licensing friction for anyone building agents or submitting jobs. On-chain coordination primitives (Rankify contracts v1) are MIT as public goods. The intent: MIT primitives spread the standard, BSL protects the orchestration layer during the critical growth phase, AGPL ensures long-term openness.