LOCAL ENGINE / 001
01 ADIL EDGE · LIVING INTELLIGENCE
LOCAL EXECUTION. LIMITLESS POSSIBILITY.
Intelligence,
closer to you.
Local-first AI for applications, devices and developer-controlled workflows.
02 Model families
Three sizes. One family.
Choose the size, then choose where it runs. Each model shows only what has been measured, with its status beside it.
-
Adil Eco
In release qualification
867 MBprocess footprint · one Apple M1 · 4K context
Model file: 484 MB on disk (Q4_K_M).
The compact model for local execution, structured tasks and developer-controlled actions on everyday CPUs.
-
Adil Core 2B
In development
2.12 GBpeak process RAM · preliminary · remote CPU host
Model file size not published.
The mid-size family member for personal AI, being measured and shaped. Every figure here is preliminary.
-
Adil Boost 4B
In development
4.80 GBpeak process RAM · preliminary · remote CPU host
Model file size not published.
The largest planned family member, aimed at personal and business workflows. Every figure here is preliminary.
Status is shown on each card. Figures are measured on different hardware, so compare them with care.
Release gates remain open. Not available for download.
03 How local intelligence works
The model proposes. Your application decides.
A conceptual sequence for the Thynkr-developed runtime environment. It describes the intended design, not a feature that has passed qualification.
Local request
Your application sends a request to the local model. Nothing leaves the device unless your code sends it.
Model inference
The local model reads the request and produces a response or a proposed tool call.
Permission check
Your host validates every proposal against the permissions and confirmations you configured.
Approved tool action
Only an approved action runs, in your code, under your control. Anything else is refused.
04 Platforms and languages
Honest about where it runs.
Platforms
- macOSApple SiliconInitial development target
- Linuxx86-64Qualification pending
- Windowsx86-64Qualification pending
- Androidarm64Qualification pending
Developer languages
- Pythoncandidate interface, not releasedCandidate · unreleased
- TypeScript / Node.jscandidate interface, not releasedCandidate · unreleased
- C / C++candidate interface, not releasedCandidate · unreleased
Platform names are text identifiers, not endorsements. The developer interface is a candidate contract and may change. No SDK package is publicly released.
05 Engineering evidence
Measured, labelled, bounded.
Every figure names what was measured and where. None is a minimum requirement or a performance promise.
Memory use, by model
These figures measure different things. Read the columns before comparing.
| Model | Model file on disk | Runtime memory | Hardware | Context | Quantisation | Status |
|---|---|---|---|---|---|---|
| Adil Eco | 484 MBQ4_K_M | 867 MBprocess footprint | One Apple M1 | 4K | Not stated | Measured |
| Adil Core 2B | Not published | 2.12 GBpeak process RAM | Remote CPU host | 8K | Q4 | Preliminary |
| Adil Boost 4B | Not published | 4.80 GBpeak process RAM | Remote CPU host | 8K | Q4 | Preliminary |
Peak RSS from getrusage on one host after a 128-token prompt and a 16-token decode. File size is not memory use. Device requirements and limits.
06 Developer experience
Built so the host stays in charge.
A conceptual walkthrough. The SDK is not released, and nothing below is runnable or downloadable.
- Your applicationOwns the data, the UI and every side effect.
- Adil SDKCandidate interfaces for C/C++, Python and TypeScript/Node.
- Thynkr-developed runtime environment and modelRuns on the device.
- Permission and confirmation policyYour rules decide which proposed tool calls ever run.
// concept only, not runnable API syntax
host.load(approvedRuntime, verifiedModel)
host.configure(context: 4K)
host.registerTools(boundedTools, confirmationPolicy)
reply = host.generate(request)
// the model returns a proposal, never an action
if (policy.approves(reply.toolCall))
host.run(reply.toolCall)
else host.refuse(reply.toolCall)Conceptual host flow. Consult the qualified SDK reference once it is published. Tool permissions and security.
07 Release journey
Gates, not dates.
Each step opens only when its evidence is in. We do not publish launch dates.
Candidate frozen
Eco A3b weights are frozen as the release candidate.
CompleteSecurity remediation
A scoped security remediation has passed independent review. Additional release qualification remains open.
In progressPlatform qualification
macOS, Linux, Windows and Android packages are not yet qualified.
PendingDeveloper availability
No downloads or SDK packages until independent acceptance and final approval.
Not startedCore and Boost
Measured and shaped in development. No release dates are published.
In development
Recent engineering milestones
- Developer interface 2.0 and engine, host and SDK 0.4.0 built for independent verification Adil Eco · SDK
- Independent security review rounds on the developer interface Adil Eco · Security
- Preliminary 8K Q4 CPU memory measurement: 2.12 GB Adil Core · Development
Newsletter
The next chapter of local AI.
Follow Adil Eco, Core and Boost. Get important engineering milestones, release announcements and developer news.
Subscriptions not open yet
The email newsletter opens after double opt-in, the newsletter privacy notice and email delivery have been verified. This page does not collect email addresses.
- Eco
- Core
- Boost
- General news
Until then: News and announcements · Changelog · Release Centre · Privacy Notice
08 Begin
Start with the evidence.
Explore the candidate model, read how the Thynkr-developed runtime environment keeps your application in charge, and follow the release gates as they open.