Releases
Releases
Every version of Debussy: what changed, how it scores, and what didn't work.
Debussy 0.1.0 Latest
The first release of Debussy, an open System 1 decision model for AI agents: safety judgements, tool-call routing and knowledge questions with an honest “I don't know”, each in a single pass, built on a 4-core CPU.
Variants
| Variant | Base model(s) | RAM | Estimated index |
|---|---|---|---|
| Debussy Duo (recommended) | Qwen3.5-4B for tool calls, K2-Horizon-7B for safety and knowledge | ~9 GB | ≈ 63.7 |
| Debussy K2 | K2-Horizon-7B, thinking switched off | ~6 GB | ≈ 62.8 |
| Debussy Light | Qwen3.5-4B | ~3 GB | ≈ 61.4 |
Results
| Benchmark | Duo | Light | Test set |
|---|---|---|---|
| When2Call (accuracy) | 93.2% | 93.2% | 600 test questions |
| WildGuardTest (average F1) | 0.936 | 0.905 | the same 500 annotator-agreed rows |
| AA-Omniscience stand-in | 53.1 | 51.8 | 266 public questions |
| HLE-MultiChoice (accuracy) | 15.0% | 10.5% | all 513 (chance 16.4%) |
Our estimates, not official Artificial Analysis scores. WildGuardTest uses the annotator-agreed rows, so it is not comparable with guard-model results on the full test set.
What's in it
- Package:
DebussyandDebussyDuo, plus an HTTP API server (python -m debussy.server) with the same endpoints as the hosted API. - Calibrated heads for both base models: safety heads fitted on 718 calibration rows, a tool-call head, an “I don't know” threshold, and a letter-bias correction.
- Fair letters: K2-Horizon leans towards some answer letters; correcting it, as chosen on practice questions only, took HLE-MultiChoice from 13.6% to 15.0%.
- debussylabs.com: model page, news, API docs and privacy notice. Sign-in and API keys are built and switched off until the hosted API opens.
Checked
- The packaged model makes the same decisions as the research pipeline (K2-Horizon: 5/5 tool-call and 15/15 safety decisions, largest probability difference 0.0000).
- Every setting was chosen on training or practice data only; each test set was used once, to report.
What didn't work
- Probes on K2-Horizon's internal state, for knowledge (51.8 vs 53.1) and for safety (0.927 vs 0.936 on test), did not beat the simpler rules.
- Combining the two models' knowledge answers, and Wikipedia retrieval for obscure questions.
- Both variants are below chance on HLE-MultiChoice.
Known limitations
- About 13 s per safety judgement on a 4-core CPU; GPU and laptop speeds not yet measured.
- English only; the safety judge reads the first 600 tokens of the request and of the reply.
- One signal for a safety system, not a complete one.