DEVELOPER SYSTEM / PUBLIC BETA

PUBLIC API ONLINE

ONE EVIDENCE CONTRACT · EVERY INTERFACE

Build against
what is known.

Discover exact artifacts, calculate a lower-bound fit, plan the missing proof, measure an endpoint, and keep the evidence class attached from terminal to agent.

Explore DeployBench Open GitHub

Install truth: `@moemodels/*` packages are not published to npm. Clone the public GitHub repository and build the workspace source.

MOEMODELS / INTERFACE BUS CORE 0.1

CLILOCAL

APIPUBLIC

SDKSOURCE

MCPSTDIO

SHARED KERNELExact artifact
Evidence class
Versioned mathUNKNOWN IS VALID

Public web + APIAvailable

CLI · Bench · SDK · MCPSource only

Passport verifierLive / local

npm packagesNot published

GitHub repositoryPublic beta

Normalized controlled runs0

INTERFACE LAYER

One system.
Seven ways in.

Each surface preserves the same distinction between sourced facts, measured evidence, calculated estimates, and unknowns. Choose the interface that fits the workflow—not a different truth model.

01SOURCE PREVIEW

CLI

@moemodels/cli · 0.1.0

Inspect exact artifacts, calculate checkpoint residency, build validation plans, and stage evaluation evidence from the terminal.

BOUNDARYRuns from a repository checkout. It does not download weights, execute a model, or certify runtime performance.

$ npm run moemodels -- plan kimi-k3 h200 --devices 8 --json

02EXPERIMENTAL

DeployBench

@moemodels/bench · 0.1.0

Send a bounded workload to an OpenAI-compatible endpoint and capture request-level TTFT, total latency, success, token usage, and throughput.

BOUNDARYA result is measured but comparison-ineligible until exact artifact, runtime, hardware, topology, repetitions, and raw evidence pass review.

$ npm run build:bench

03LOCAL VERIFIER

Deployment Passport

@moemodels/bench/browser · 0.2 CONTRACT

Pack compatible DeployBench trials under one content digest, recompute their summaries, inspect reproducibility gates, and verify optional operator signatures.

BOUNDARYIntegrity proves that bytes match. A self-signature proves control of a key. Neither proves the declared deployment, admits a run, or makes it comparison-eligible.

$ node packages/bench/dist/cli.js verify passport.json --json

04LIVE

HTTP API

PUBLIC JSON · V1

Query the versioned registry, calculated fit, deployment-validation plans, and evaluation evidence over public JSON routes.

BOUNDARYFit and plan outputs are deterministic calculations—not measured performance, runtime certification, or procurement advice.

$ curl https://moemodels.ai/api/v1/registry

05SOURCE PREVIEW

TypeScript SDK

@moemodels/sdk · 0.1.0

Use one typed, browser-safe client for registry, fit, evaluation, and planning endpoints without duplicating query serialization.

BOUNDARYThe package is not published to npm. Import it from the GitHub source workspace while its public contract is hardened.

$ npm run build:sdk

06LOCAL / READ-ONLY

MCP server

@moemodels/mcp · 0.1.0

Give compatible coding agents structured access to model search, sourced comparisons, static fit, evidence summaries, and validation-plan exports.

BOUNDARYThe local stdio server cannot run benchmarks, mutate the registry, fetch remote content, or present reported claims as measurements.

$ npm run build:mcp && node packages/mcp/dist/index.js

07PUBLIC BETA

GitHub

MONOREPO · APACHE-2.0 CODE

The schemas, registries, calculation engine, adapters, application, packages, fixtures, and contribution checks live in one source repository.

BOUNDARYClone and build from GitHub source while packages remain unpublished. Third-party weights, datasets, benchmark content, and marks retain their own rights.

PUBLIC API / V1

Evidence over HTTP.
No client required.

The current API is public and read-only. It exposes sourced registry records, deterministic calculations, and evaluation evidence. A stability guarantee has not been published yet.

Open the full contract

METHODENDPOINTRETURNS

GET/api/v1/registryExact model, hardware, compatibility, and provenance records GET/api/v1/fit?model=kimi-k3&hardware=h200&devices=8Calculated static checkpoint-residency lower bound GET/api/v1/plan?model=kimi-k3&hardware=h200&runtime=vllmValidation plan with ordered evidence gates and explicit unknowns GET/api/v1/evaluations?model=qwen3-30b-a3bReported claims and normalized runs without evidence-class merging GET/api/v1/openapiMachine-readable OpenAPI 3.1 contract

GITHUB SOURCE

From clone
to decision.

Until packages are published, the supported developer path is the public GitHub monorepo and lockfile. These commands build every current package and generate a validation plan; they do not execute a model.

Understand the open-core boundary Measure an endpoint with DeployBench

moemodels / quickstartNODE 22.13+

# 01 · clone the public source
git clone https://github.com/SamSnead85/moemodels.git
cd moemodels

# 02 · restore the committed workspace
npm install
npm run build:packages

# 03 · inspect exact artifacts
npm run moemodels -- models

# 04 · export a deterministic validation plan
npm run moemodels -- plan moonshotai/kimi-k3 \
  nvidia/h200-sxm-141gb --devices 8 \
  --runtime vllm --json

MEASUREMENT SURFACE

Ready to produce
evidence?

Start with DeployBench when you control an OpenAI-compatible endpoint. Start with the Workbench when the deployment question still needs to be defined.

Open DeployBench Build a plan

Build againstwhat is known.

One system.Seven ways in.

CLI