SmartNETWORK One presents

SmartROOT One

When AI goes above and beyond.

One model. Every modality. Your weights. A single transformer — one set of weights, one vocabulary — that chats, calls tools, sees, listens, and creates images, video, music and speech. Behind an OpenAI-compatible API.

Get started View on GitHub

1weight set

8modalities

37,120unified tokens

139.7Mparameters

The model

Not a wrapper. A real model.

SmartROOT One is a from-scratch multimodal AI model — not a gateway, proxy, router or wrapper around other models. Every capability lives in the same weights. No other model in the request path. Ever.

◈

One weight set

Chat, tools, vision, audio and generation all run from a single decoder-only transformer. One artifact, one deploy, zero model-zoo sprawl.

⬡

Unified token space

Text, image codes and audio codes share one vocabulary. The model learns every modality the same way it learns language — by predicting tokens.

⛨

Commercially-safe training

Trained exclusively on permissively licensed data — ODC-By, CC-BY, CC0, Apache-2.0, MIT. Clean data means weights you can actually ship in a product.

⇌

OpenAI-compatible API

Chat completions, embeddings, image, video, music, speech and transcription endpoints that speak the API your tools already use — including OpenCode.

Capabilities

Every modality. One brain.

Eight endpoint families, all served by the same SmartROOT One weights.

💬

Chat

Conversation and instruction following over an OpenAI-compatible completions API, with streaming.

POST /v1/chat/completions

🛠️

Tool calling

Structured function calls the model emits itself — ready for agents and orchestration.

POST /v1/chat/completions

🧭

Embeddings

Semantic vectors from the same weights — search, clustering, retrieval.

POST /v1/embeddings

🖼️

Image generation

Text-to-image through discrete image codes emitted by the transformer itself.

POST /v1/images/generations

🎬

Video

Short video as sequences of generated frames — same codes, new dimension.

POST /v1/videos/generations

🎵

Music

Compose audio from a text prompt via the model's own audio code vocabulary.

POST /v1/music/generations

🗣️

Speech synthesis

Text-to-speech: the model speaks by generating its own audio tokens.

POST /v1/audio/speech

📝

Transcription

Speech-to-text in the same token stream — the model hears and writes.

POST /v1/audio/transcriptions

Architecture

How one model does everything

One vocabulary

32,000 BPE text tokens + 4,096 image codes + 1,024 audio codes = one token space of 37,120. Everything the model reads or writes is a token.

Modality tokenizers

Small vector-quantized autoencoders turn pixels and waveforms into discrete codes — the image/audio analogue of BPE. They ship inside the model bundle.

A single transformer

One decoder-only transformer predicts the next token — whether that token is a word, an image code or an audio code. Generation and understanding share weights.

Chained training

Pretrain on curated public text → fine-tune on instructions and tool calls → unify modalities with an anti-forgetting text mix. Every stage continues from the last.

text

→

image codes

→

SmartROOT One transformer

→

words · images · video · music · speech

Quickstart

Speak to it in your API

The server is OpenAI-compatible. Point any SDK — or OpenCode — at it and go.

curl http://localhost:8080/v1/chat/completions \
  -H 'Content-Type: application/json' \
  -d '{
    "model": "smartroot-one",
    "messages": [{"role": "user", "content": "Say hello."}]
  }'

from openai import OpenAI

client = OpenAI(base_url="http://localhost:8080/v1",
                api_key="smartroot-local")

resp = client.chat.completions.create(
    model="smartroot-one",
    messages=[{"role": "user", "content": "Say hello."}])
print(resp.choices[0].message.content)

// opencode.jsonc
{
  "providers": {
    "smartroot": {
      "npm": "@ai-sdk/openai-compatible",
      "name": "SmartNETWORK One",
      "options": {
        "baseURL": "http://127.0.0.1:8080/v1",
        "apiKey": "smartroot-local"
      },
      "models": { "smartroot-one": { "name": "SmartROOT One" } }
    }
  }
}

// then: /models → SmartNETWORK One → SmartROOT One

Every endpoint, one server

Chat, embeddings, images, video, music, TTS and STT — all served from one running process and one weight set.

Own the weights

The full stack — training, codecs, server — is open on GitHub. Train it, tune it, ship it. It's yours.

Clone the repo

When AI goes above and beyond.

One model. Every modality. Your weights.

GitHub Quickstart