Introducing Debussy
An open System 1 decision model for AI agents: fast decisions and honest answers, on an ordinary CPU.
Updated the same day with the results of the last two experiments: the letter-bias correction (HLE 13.6% → 15.0%, index ≈ 63.4 → ≈ 63.7) and a safety probe that didn't hold up.
Most of what an AI agent does all day isn't writing essays. It's making small decisions. Is this request safe to fulfil? Should I call this tool now, or first ask the user for the missing date? Do I actually know the answer, or am I about to make one up?
Today we're releasing Debussy 0.1, an open model built for exactly those calls. It is small, it runs on a CPU, and when it isn't sure, it says so.
A System 1 model
Psychologists describe two modes of thinking: a fast, intuitive System 1 and a slow, deliberate System 2. Most recent AI progress has gone into System 2, with models that reason step by step before answering. Debussy goes the other way on purpose.
Debussy reads its input once and answers a few short yes/no questions in a single pass. It never writes out its reasoning or generates text. Instead it returns a probability for every option, the decision-model format used by the Artificial Analysis Decision Index. Your code gets a number it can act on: set a threshold, route the hard cases to a person, or log the confidence.
- Safety judge. Is the user's request harmful? Did the assistant refuse? Is the reply itself harmful?
- Tool-call router. Call the tool, ask the user for missing details, or say it can't be done.
- Honest knowledge. Pick the right option, or answer “I don't know”.
Results
We measured Debussy on the four parts of the Decision Index, using test questions that were never used to fit or choose anything.
| Benchmark | Debussy Duo | Debussy Light |
|---|---|---|
| When2Call (tool calls, accuracy) | 93.2% | 93.2% |
| WildGuardTest (safety, average F1) | 0.936 | 0.905 |
| AA-Omniscience stand-in (knowledge) | 53.1 | 51.8 |
| HLE-MultiChoice (hard questions, accuracy) | 15.0% | 10.5% |
| Estimated Decision Index | ≈ 63.7 | ≈ 61.4 |
These are our estimates, not official scores. WildGuardTest uses the rows where all three annotators agree, as Artificial Analysis does, so it isn't comparable with guard-model results on the full test set. The Omniscience questions are private, so we used Artificial Analysis's 266 public ones. Debussy Light scores 0.899 on all 880 agreed WildGuard rows, and its index estimate uses that number.
Two specialists beat one generalist
The biggest surprise of the project was that different small models are good at different decisions. We ran the same questions, with the same training and test rows, on two open models:
- Qwen3.5-4B is the better tool-caller: 94.0% against 89.3% on the same 300 When2Call questions.
- K2-Horizon-7B, a reasoning model that nobody else uses for this, is the sharper judge when run with its thinking switched off: 0.936 against 0.905 on the same 500 safety rows. It also knows its limits better on knowledge questions.
So the recommended Debussy Duo sends tool-call decisions to Qwen and safety and knowledge decisions to K2-Horizon. Each decision is still a single pass. If you only have about 3 GB of memory to spare, Debussy Light runs on Qwen alone.
Honest by design
On knowledge questions, the “I don't know” option is never shown to the model. Debussy answers only when its confidence in the best option clears a threshold set on a separate practice set, and abstains otherwise. On Artificial Analysis's obscure public questions, K2-Horizon answers only 9% of them, and when it doesn't know the answer it says “I don't know” 98% of the time instead of guessing wrong.
Fair letters
Small models lean towards some answer letters whatever the question. K2-Horizon favours “B” and avoids the last option. We measured that lean on practice questions where every letter is equally likely to be right, and Debussy now corrects for it. On HLE-MultiChoice, applied once, that took K2-Horizon from 13.6% to 15.0%: still below chance, but honestly measured.
Built on a 4-core CPU
Debussy was built and evaluated on a 4-core cloud CPU with 15 GB of memory: no GPU, no API key, no paid services. It doesn't change the base models' weights. What it adds is the questions it asks and a few hundred calibrated numbers, fitted on public training data.
What didn't work
We think the misses are as useful as the hits, so here they are:
- A probe on the model's internal state to predict when an answer would be right scored worse than the model's plain confidence (51.8 against 53.1). We kept the simple rule.
- Our own check caught us. With random noise in place of real data, the best of 12 probe settings still “won” ordinary cross-validation. Since then, every new idea is judged with nested cross-validation.
- A second probe, for the safety judge, looked better than the calibrated heads inside the calibration data but not on the 500 test rows (average F1 0.927 against 0.936), and it would have doubled the time per judgement. We kept the heads.
- Combining the two models' knowledge answers and Wikipedia retrieval for obscure questions gave little or nothing.
- Hard multiple choice: both variants score below chance (16.4%) on HLE-MultiChoice. A System 1 model does not solve these questions.
Limitations
- Speed on small CPUs: about 13 s per safety judgement on the machine it was built on. GPUs and modern laptops should be much faster; we haven't measured them yet.
- English only, and the safety judge reads only the first 600 tokens of the request and of the reply.
- Not a complete safety system. Use it as one signal, with human review for anything high-stakes.
What's next
- An independent evaluation, so the numbers above stop being our estimates.
- Speed measurements on GPUs and laptops, and a comparison with established guard models on the full WildGuard test set.
- A Hugging Face release and a hosted demo you can try in the browser.
- Better speed on small machines, so a safety judgement doesn't need 13 seconds.
Try Debussy. The code, model card and setup instructions are on GitHub under Apache-2.0.
Thanks
Debussy stands on open work: the Qwen team's Qwen3.5-4B, the Institute of Foundation Models' K2-Horizon-7B, llama.cpp, and the When2Call, WildGuardMix, MMLU-Pro, Humanity's Last Exam and AA-Omniscience datasets.
Debussy was built by Debussy Labs with help from Claude, Anthropic's AI assistant. It is not an Anthropic model. The name is a nod to that: it's after the composer Claude Debussy, and the logo is a Clair de lune crescent and an Arabesque curl.