LOCAL FIRST · TASK ROUTING · QUALITY GATES

Not one "best" model.
The best route for the task.

Digi Anton separates routine orchestration, substantive analysis, code, knowledge retrieval and visual production. The model is chosen by work type, required depth, cost and verification result.

Principle

Local models handle the bulk of work without sacrificing quality

Simple classification, extraction, drafting and tool calls don't need an expensive cloud model. Complex analytical reports, long context or final review get a stronger route. If the result doesn't pass a domain quality gate, the task is escalated.

Role map

Six specialised routes

ROUTINE · LOCAL

Qwen 3.5 35B

Tools, classification, extraction, short drafts, recurring tasks and fallback routing.

ANALYSIS · LOCAL

DeepSeek V4 Flash

Substantive analysis, long context, research, book handling and complex drafts.

CODE · LOCAL

Qwen Coder

Routine coding tasks and technical scaffolding; system implementation and evolution stay under Codex control.

ENGINEERING

Codex

Architecture, diagnostics, OpenClaw changes, testing, decision memory and promotion of verified improvements.

KNOWLEDGE RETRIEVAL

BGE-M3 and retrieval

Vector search and knowledge layer return relevant context before calling the generative model.

AMPLIFICATION · CLOUD

DeepSeek Pro and Kimi

Important results that didn't pass the local route check, complex and high-value final tasks.

Selection logic

How the model is chosen

01 · CLASSIFICATION

Task type

Routine, analysis, code, retrieval, visual or external action.

02 · ROUTING

Minimum sufficient route

Start with the free local model if its quality matches the task.

03 · VERIFICATION

Domain quality gate

Facts, sources, completeness, tests, format and practical usefulness.

04 · AMPLIFICATION

Amplification when needed

A stronger local or limited cloud model receives the problem stage.

Practical rules

Cost should not become hidden quality reduction

Routine — local — repeated tasks don't default to a paid model.

Quality over savings — important results get amplified when the local answer is weak.

Knowledge first — own knowledge is extracted before forming an answer.

Dedicated resources — images and video go to their intended resources.

Controlled amplification — paid routes are explicit, bounded and tied to a specific task.

Timeliness

Versions change, roles persist

This page documents a qualified architecture as of August 2026. Specific builds may be updated after testing, but the split across routine, analysis, code, knowledge retrieval, visual tasks and cloud escalation remains a project principle.