Benchmarks · All Adil Edge models
Every result with its conditions and its limits.
This is the permanent index of benchmark reports for Adil Edge models, including models that are planned but not yet trained. Each report states how it was measured, what it can and cannot support, and the release status of the model. No model described here is released.
Reports
Comparative engineering benchmark of the Adil Eco V1.1 release candidate: knowledge, maths, coding, instruction following, structured JSON, CPU memory and context latency. Internally conducted; not independent third-party certification.
Measured on RunPod cloud servers; published .
No benchmark exists: Adil Core is planned and training has not begun. Results will be published here only after training, evaluation and measurement.
No benchmark exists: Adil Boost is planned and training has not begun. Results will be published here only after training, evaluation and measurement.
Later Adil Eco releases will be added as separate reports; earlier reports stay published.
1. Internal development evaluations
Internal · not published as claims
Engineers score candidate models on internal development test sets to choose between them. Because candidates are selected on these sets, their scores are optimistic, and we do not publish them as performance claims.
2. Sealed acceptance
Gate failed
Before any release claim, a candidate is scored once on a sealed test set that engineers have not seen, against a gate that is locked before the run.
For the current Eco candidate, the locked gate required a difference of at least 5 percentage points over the base model. The measured difference was 4.16 percentage points, so the gate failed and no advantage over the base model is claimed.
That sealed test set has been used and will not be reused.
3. Release acceptance
Prepared · not run
A fresh sealed acceptance evaluation, with a new test set recorded before use, has been prepared but has not been run. Until a candidate passes it and the other release conditions are met, Adil Eco will not be released.
4. CPU and memory measurements
The current Adil Eco candidate was measured on a single Apple M1 (8 GB) test host with 4 threads. These are interim engineering measurements, not benchmarks and not system requirements.
Not comparable: the engineering benchmark reports process memory (RSS) on a RunPod cloud server, a different metric on a different host.
Method rules