IDO-1

IDO-1: the brain behind IDOLL.

Our in-house large language model stack, tuned for English conversation, long-term interaction, and real-time robotics.

What IDO-1 is

Built for robots, not just for chat tabs.

IDO-1 is a family of models we operate and fine-tune specifically for IDOLL and future robots. It’s built on strong open foundations, then trained and adjusted for:

Human-like conversation in English
Stable long-term memory behavior
Low-latency generation so physical motion feels natural

Benchmarks

Benchmarked on real tasks, not just vibes.

We evaluate IDO-1 using widely recognized benchmarks so we can track how it performs and where it needs to improve. Below are some of the tests we run regularly.

MT-Bench – multi-turn conversation quality.

MT-Bench scores models on how well they handle multi-turn dialogue across different topics. IDO-1 is evaluated against standard baselines so we can see how natural, helpful, and consistent its responses are over longer conversations.

MT-Bench comparison: IDO-1 vs widely used open and proprietary models. (Higher is better; exact scores shown in the chart.)

MMLU – broad knowledge and reasoning.

MMLU tests a model’s knowledge and reasoning across dozens of domains – from humanities to STEM. We use it to ensure IDO-1 can handle everyday questions and explanations without collapsing the moment you go beyond small talk.

MMLU performance across domains for IDO-1. This helps us ensure that ‘Can you explain this to me?’ is a safe question to ask IDOLL in everyday use.

Task-specific and internal evaluations.

In addition to public benchmarks, we run our own evaluations focused on how IDO-1 behaves inside a robot: robustness in noisy audio environments, consistency between speech and proposed motion, and stability over long sessions where the user comes back again and again.

Internal evaluations for interaction stability and behavior quality over time.

Why IDO-1

Why not just use GPT-5 or another API?

External APIs are amazing, and we still use them where they make sense. But if your product is a physical robot, you can’t outsource its entire brain forever.

Latency and cost need to be under our control
Behavior needs to be tuned for motion, not just text
We need to guarantee stability even if external models change overnight
We want to offer a device that can become more independent over time

Running IDO-1 on our own infrastructure lets us design IDOLL’s experience end-to-end instead of stitching together someone else’s roadmap.

Real-time behavior

Designed to respond before a moment feels awkward.

In a physical interaction, a 4–5 second pause already feels wrong. IDO-1 is tuned for streaming responses, so IDOLL can start reacting quickly – turning its head, making a small movement, or starting to speak while the rest of the answer is still being generated.

If you want, you can add a simple bar chart later showing average time to first token and average time to full response — but you don’t have to claim specific milliseconds here.

Continuous improvement

The IDO-1 you meet now won’t be the last version.

We treat IDO-1 as a living system. New training runs, new safety layers, and new interaction data from real users all feed back into it. IDOLL is designed so its brain can be upgraded over time without you having to buy a new robot every year.