Performance & research
Performance you can verify.
Faster is useful. Measurable is essential.
Adil Edge's performance roadmap focuses on the full experience of local AI: response latency, memory use, quality, energy efficiency, startup time and workload predictability.
An early prototype observation
Initial measurement, clearly labelled.
- Decode steps per second
- ≈ 72.6
- Reported peak process RSS
- ≈ 1.44 GB
- Measured decode steps
- 13
During an early supervised CPU-only runtime observation on an Apple M1 test host, a short request produced approximately 72.6 decode steps per second and a reported peak process RSS of approximately 1.44 GB. No swap growth was observed during that specific run.
Measurement limits: This was a single short instrumentation check with 13 measured decode steps, not an independently repeated benchmark or a certified result for Micro, Core or VPS. End-to-end generated-token throughput and decode-step throughput are different measurements. These figures should not be compared directly with vendor model benchmarks.
Research targets
What we aim to improve.
Research objectives — not achieved results
The performance research programme has proposed objectives including:
- At least 20% improvement in CPU decode throughput in suitable matched comparisons
- 20–30% lower peak memory consumption for comparable workloads where feasible
- At least 25% lower cold-start latency as an exploratory objective
- Improved energy efficiency, measured in useful work per joule
- No material loss of output quality in accepted target tasks
- Reproducible results using identical hardware and evaluation conditions
These are research objectives only; none is a published product-level performance achievement. Results may vary by model, processor, quantisation, workload and operating system.
Memory research
Less overhead. More useful work.
Our ongoing investigations examine how compact model weights are loaded, represented and retained in memory. We will evaluate configuration-level approaches and carefully controlled changes before making any memory-reduction claims.
A previously observed memory allocation should not be presented as guaranteed reclaimable RAM: process RSS, anonymous allocations and file-backed cache can behave differently.
Performance modes roadmap
Adaptive performance, only when validated.
Research modes under consideration include:
- Eco:
- Lower resource demand where practical.
- Balanced:
- Sensible everyday operating configuration.
- CPU Boost:
- Higher throughput on capable processors.
- Hardware Turbo:
- Optional hardware acceleration when it is available and justified.
Research roadmap These names describe a future research roadmap, not current selectable product settings.
Benchmark integrity
Our intended approach includes defined workloads, versioned fixtures, immutable evidence, correct measurement units, independent review and publication of meaningful limitations.