Apple Silicon

A laptop on a measurement platform beside test instruments and a desktop computer, representing a proof of concept.

Edge0 Part 5: M2 Air 16GB Measurements and an Internal PoC

I ran an agent that connects Microsoft Learn search, Edge0 streaming, and sentence-by-sentence Korean translation on a 16GB M2 MacBook Air. The sample Gist, checked on September 13, 2026, includes the complete code and an execution result with a TTFT of 53.47 seconds. This final article covers the test question, time through search and first response, evidence behind the answer, interpretation of output rate, follow-up optimization, and conditions for an internal PoC. The results use the client-side measurements described in Part 4. ...

Blue glass panels following an initial gap along a timing rail between a timing instrument and a laptop.

Edge0 Part 4: Measuring TTFT and Streaming Output Rate

As of September 13, 2026, the Python sample for this series records first-response latency and output rate while receiving Edge0 SSE responses. This article connects the grounding prompt from Part 3 to an actual generation request. The article covers measurement boundaries, progress display, stream reception in a worker thread, sentence-by-sentence Korean translation, and interpretation of the results. In my 16GB M2 MacBook Air experiment, the first response took tens of seconds, so I recorded the time before and after it separately. ...

A magnifying glass over highlighted document passages and a glass box holding three selected reference cards.

Edge0 Part 3: A Local RAG Agent with Microsoft Learn Search

As of September 13, 2026, the sample Gist provides Python code that connects the Edge0 server, Microsoft Learn CLI, and Apple Translation. Building on the environment from Part 2, this article explains how retrieved material becomes model input. The article covers the division of work between host and model, question normalization and translation, search-result selection, grounding prompts, and conversation context. The host determines the search sequence; the model reads the supplied material and constructs an answer. ...

A compute device, document cards, and translucent speech bubbles connected in front of a laptop.

Edge0 Part 2: Microsoft Learn and Apple Translation Setup

As of September 13, 2026, the sample in this series uses Edge0-35B, Microsoft Learn CLI, and the macOS Translation framework. Following the responsibilities established in Part 1, the inference server and Python agent host run in separate terminals. This article covers installing Edge0 and its model, starting the streaming server, searching official documentation, translating between Korean and English, and normalizing questions. The complete files are README.md, apple_translate.swift, and microsoft_expert_agent.py in the sample Gist. ...

Glass blocks above a silver laptop and stacked storage devices representing local model inference.

Edge0 Part 1: A 35B Local LLM and Grounded Agent on a 16GB Mac

As of September 13, 2026, Edge0 provides an MLX backend for Apple Silicon and the Edge0-35B-A3B-preview model. I ran the model on a 16GB M2 MacBook Air, then built a technical expert agent that answers using Microsoft Learn references. This series covers the requirements for running a 35B-class model, changes in answers after adding references, Korean input and output, streaming measurements, and conditions for an internal PoC. The complete runnable code is available in the Edge0 Microsoft Technology Expert Agent Gist. ...