ADIL ECO · REQUIREMENTS AND MEASUREMENTS
What Eco needs, and what has been measured.
Development measurements for the Adil Eco V1.1 A3b candidate, with the limits of what they show. Neither the Eco V1.1 A3b model nor the SDK is publicly released: there is no download, package or installer, and nothing on this page announces one.
Overview
Adil Eco is the compact tier of the Adil Edge family. The A3b candidate is a post-trained Qwen3-0.6B derivative (Apache-2.0 foundation) distributed as a 4-bit (Q4_K_M) file of about 484 MB, run on the CPU by the Adil Native Runtime Environment, which is built on llama.cpp, which is MIT-licensed.
It is intended for bounded text tasks, structured output and host-controlled tool requests. It is a small model: it can misunderstand a request, invent arguments, misreport numbers it copies from tool results and follow instructions embedded in tool results. Treat every output as untrusted until your application has checked it.
Hardware and operating requirements
| Current position | |
|---|---|
| Processor and system | Apple silicon Mac (M1 or later). CPU inference only; no GPU is used or required. |
| Memory | The reference test host was an Apple M1 with 8 GB of memory. On that Apple M1 host the peak process footprint was 880–900 MiB at a 4K context and 1,338–1,352 MiB at 8K, and resident memory reached about 1.16 GB at 4K and 1.61 GB at 8K. These engineer-measured figures come from an earlier build and have not been re-measured on the current candidate. They are measurements on one Apple M1 host with 8 GB; other hosts and memory sizes are unverified, and they are not minimum or recommended requirements. |
| Disk | The model file is 484,219,776 bytes (about 484 MB), plus the engine libraries. |
| Operating system | Tested on macOS 13.4.1 only. The candidate binaries are built with macOS 13.0 as a build target; that is not a tested or supported minimum, and other macOS versions are untested. |
| Context | 4K is the default development profile and 8K the maximum. The engine does not accept longer contexts. |
| Network | None for inference (see below). A network is needed only to obtain software in the first place. |
Intel Macs, Linux, Windows, Android and web browsers are not supported by the current candidate package. Qualification for Linux x86-64, Windows x64 and Android ARM64 is pending; see platform status below.
Offline and local inference
Inference runs on the machine that runs your application. The Adil Native Runtime Environment is started in offline mode (no model downloads), the SDKs talk only to a local service, and nothing in the SDK sends telemetry. In engineering tests, a subset of the regression suite (the packaged C example, the Python native-adapter tests and the TypeScript real-host tests) was run inside a macOS sandbox that denies network access, with a control showing that a network connection was refused.
Local does not mean your whole application is disconnected: tools you register, updates, accounts and any remote services you add can still use the network, and your application controls those.
Performance and memory measurements
Historical development measurements on one Apple M1 (8 GB, macOS 13.4.1, four CPU threads, CPU-only build, frozen A3b Q4_K_M, fresh process per row, three repetitions):
| Context | Decode tokens/s | First token (s) | Peak footprint (MiB) |
|---|---|---|---|
| 4K | 21.8–22.1 | 12.3–13.8 | 880–900 |
| 8K | 9.9–10.2 | 37.8–41.9 | 1,338–1,352 |
The first-token figures are for a prompt that nearly fills the context window, so they describe a worst case, not a typical short request. Peak footprint and resident memory at the end are different metrics; neither is a universal RAM requirement. These measurements precede later security corrections and do not qualify the current SDK candidate.
Supported platforms and qualification limitations
| Platform | Status |
|---|---|
| macOS, Apple silicon | Initial development target. Tested on one Apple M1; not qualified for release. |
| Linux x86-64 | Qualification pending |
| Windows x64 | Qualification pending |
| Android ARM64 | Qualification pending |
A platform is described as qualified only after independent acceptance of the model, the engine, the SDK and the package on that platform. Until then, a result on one machine is evidence, not a support statement.
What the quality evidence shows
Independent evaluation of the A3b candidate returned mixed results: some structured-output results are encouraging, a predefined sealed overall-improvement threshold was not reached, and one mathematics evaluation was weaker than the foundation model. Held-out evaluation also found that the model can misreport numbers copied from tool results, drop a stated unit, fill a wrong email recipient or swap a currency pair. Only host-side confirmation stops the resulting actions.