trio.ai — Open Source Local AI Models from RioCloud

Most teams discover the same wall at the same time: their OpenAI bill crosses the cost of a small server, a compliance review flags customer data going out to a third-party API, or a flaky inference endpoint takes down a feature in production. trio.ai exists for that moment. It is a family of open weights, a tiny Python package, and a permissive licence — the parts you need to move AI from "rented from a vendor" to "owned by your team". See the trio.ai page for downloads and benchmarks, or skim the rest of this article to understand which model fits your workload.

What is trio.ai?

trio.ai is an open-source language-model project published by RioCloud Solutions. It ships as six instruction-tuned models covering the practical size spectrum — from a 1 GB model that runs on a Raspberry Pi to an 18.6 GB model that competes with mid-tier hosted APIs on a single 24 GB GPU. All weights are released on HuggingFace under the trioai-org organisation, the runner code lives at github.com/iampopye/trio, and the convenience wrapper is on PyPI as triobot.

The project is built around three opinions. First, local should be the default for any workload that touches customer data, internal documents, or anything you would not paste into a public chatbot. Second, small models are good enough for most production tasks — classification, extraction, summarisation, routing — and they are dramatically cheaper. Third, tooling matters more than size; a 4 GB model with a clean Python API and a healthy quantisation story will beat a 70 GB model that needs a research engineer to babysit it.

What models are in the trio.ai family?

The six published models split into three weight classes — nano, medium, pro — each with a base variant and an instruct variant. The instruct variants are what most teams should use; the base variants are for fine-tuning.

All six models share the same tokenizer and the same chat template, which means you can swap sizes without rewriting your prompts. Most teams develop against trio-medium-instruct on a laptop and then promote to trio-pro-instruct for production — or down to trio-nano-instruct for an embedded device. The trio.ai page lists the exact HuggingFace links and SHA256 hashes for each weight file.

When should you choose a local LLM over a cloud API?

Cloud APIs (OpenAI, Anthropic, Google) are still the right choice for plenty of work — anything that needs frontier reasoning, anything with very spiky traffic, anything where you are still figuring out the use case. Local models start to win when one of these conditions is true:

1. The data is sensitive. Health records, financial transactions, employee documents, anything covered by GDPR, DPDP, or a hospital's BAA. If your security team is uncomfortable with the data leaving your VPC, a local model is the cheapest way to make the conversation end.
2. The volume is high and the task is narrow. Classifying 5 million tickets a month at 200 tokens each will run you four figures on a frontier API; trio-medium-instruct handles that on a single L4 for the cost of the instance.
3. You need predictable latency. A local model behind your load balancer has no provider outage risk, no rate limit, and round-trip times measured in milliseconds rather than hundreds of milliseconds.
4. You want to fine-tune. You own the weights, so you can fine-tune on your own data, ship the result inside your product, and never worry about a model-deprecation email.
5. You ship at the edge. Retail kiosks, factory-floor tablets, in-car assistants, offline mobile apps — these need inference without an internet round-trip, and that is exactly what trio-nano is built for.

If none of these apply, a hosted API is probably fine. We help clients run both, often side-by-side — see our writeup on VibeMaster, which orchestrates trio.ai alongside GPT, Claude, and Gemini in a single workflow.

How do you install and run trio.ai?

The fast path is three commands. You install the Python wrapper, pull a model from HuggingFace, and call generate():

1. Install the package. Run pip install triobot in a clean virtual environment. The wheel is small (under 200 KB); it pulls in transformers, torch, and huggingface_hub as dependencies.
2. Pull a model. Run triobot pull trio-medium-instruct. The CLI downloads weights from HuggingFace to ~/.cache/trioai/, verifies the SHA256, and prints the absolute path. First-run download for trio-medium is about 4.7 GB.
3. Load and call it. In Python: from triobot import Trio; m = Trio("trio-medium-instruct"); print(m.chat("Summarise this email in three bullets: ...")). The first call warms the GPU; subsequent calls run at single-digit-second latency on a laptop, sub-second on a server GPU.
4. Serve it as an API (optional). Run triobot serve trio-medium-instruct --port 8080 to expose an OpenAI-compatible HTTP endpoint. Any client that speaks /v1/chat/completions — LangChain, LlamaIndex, OpenWebUI, your own app — works without changes.
5. Quantise for smaller machines (optional). Run triobot quantize trio-medium-instruct --bits 4 to produce a 1.4 GB int4 file that runs on an 8 GB MacBook with negligible quality loss for most tasks.

That is the whole stack. There is no separate inference engine, no Docker image to build, no Kubernetes operator to install. If you outgrow the single-process serve command, swap in vllm or text-generation-inference — the weights are standard HuggingFace format, so anything that loads a Llama or Mistral checkpoint will load trio.ai.

How does trio.ai compare to other open models?

The open-weights landscape has matured fast. trio.ai is not trying to top a leaderboard; it is trying to be the obvious default for a small team that wants a Python-first, well-packaged local model with a clean upgrade path. The closest neighbours are Llama 3.x, Mistral, Qwen, and Phi.

If you already have a working stack on Llama 3 or Mistral, there is no urgent reason to switch. trio.ai shines for teams that have never deployed a local model before and want one obvious package to install, one CLI to learn, and a sensible default at every size class.

Where does trio.ai fit in the RioCloud stack?

trio.ai is one of two open products we maintain. The other is VibeMaster, our AI agent command centre, which orchestrates GPT, Claude, Gemini, and trio.ai in the same workflow. A typical pattern: VibeMaster routes simple, high-volume requests (intent classification, language detection, sentiment tagging) to a local trio-medium-instruct instance, and reserves frontier APIs for the genuinely hard reasoning steps. Teams that adopt this pattern usually see 60–80% of inference cost move from per-token API spend to a fixed monthly server cost — without losing quality on the user-facing answer.

On the consulting side, we run a small AI engineering practice that helps teams pick the right deployment shape (cloud API vs local model vs hybrid), set up fine-tuning pipelines, and bolt the result into existing apps. If you want a structured starting point, the AI marketing use-cases guide shows how marketing teams plug local models into their stack without ripping out tools they already use.

Frequently asked questions about trio.ai

Is trio.ai free to use commercially?

Yes. trio.ai weights are released under a permissive licence that allows commercial use, redistribution, and fine-tuning. The triobot Python wrapper is Apache 2.0. See the LICENSE file in the GitHub repo for the exact terms.

Which trio.ai model should I start with?

Start with trio-medium-instruct (4.7 GB). It fits on most laptops, runs at usable speed, and covers the majority of production tasks. Drop down to trio-nano-instruct only if you are constrained to CPU or an edge device; jump up to trio-pro-instruct only when you have a measured quality gap on a specific workload.

Where can I download trio.ai weights?

All six models are on HuggingFace under the trioai-org organisation. The easiest path is triobot pull <model-name>, which handles the download, the SHA256 verification, and the cache placement. The CLI is installed by pip install triobot.

Does trio.ai support fine-tuning?

Yes. The base variants (trio-nano, trio-medium, trio-pro) are designed as fine-tuning starting points. We recommend LoRA on trio-medium for most domain tasks — it converges in a few hours on a single GPU and the resulting adapter is small enough to ship in your container.

Can I run trio.ai without a GPU?

Yes for trio-nano (CPU-only is fine) and yes for quantised trio-medium (int4 runs on an M-series Mac or a modern x86 laptop). trio-pro needs a GPU with at least 24 GB of VRAM at full precision, or 12 GB at int4.

How does trio.ai handle data privacy?

By design, nothing leaves the machine you run it on. There is no telemetry, no callback, no usage reporting. The wrapper only contacts HuggingFace during the initial pull, and after that all inference is fully local — which is the whole point of running an open model.

Can RioCloud help me deploy trio.ai in production?

Yes. We deploy trio.ai for clients across India, the UK, the UAE, and Singapore — usually as part of a wider AI build that also touches VibeMaster, RAG pipelines, and existing data systems. Book a call and we will scope a deployment for your stack.

Next steps

If you want to try trio.ai today, run pip install triobot && triobot pull trio-medium-instruct and you will have a working local model in about ten minutes. If you want help wiring it into a real application — fine-tuning, RAG, orchestration with hosted models, or a production deployment — book a 30-minute call and we will scope it together. For the wider picture of where trio.ai fits in our AI stack, read the VibeMaster overview or the guide to building chatbots that use local models under the hood.