Flint Brings Coding Agents to Your Mac. Check the Memory Before the Speed.

2026-10-22

Maestro’s take. Flint is worth testing on hardware you already own. Buying a new Mac on a vendor speed chart is a much larger bet.

TL;DR

Recursive released Flint on October 21. Its announcement describes a roughly 40-billion-parameter coding model that activates about 2.5 billion parameters per token. Recursive reports more than 125 tokens a second on an M4 Pro at short contexts. That is a vendor measurement, not our test.

The weights carry an Apache 2.0 license. The optimized rec engine currently requires an M4 Pro, M5 Pro, or M5 Max, at least 48 GB of unified memory, and about 30 GB of free disk space. It is optimized for one session.

Why Maestro users care

Local inference gives you another place to route bounded coding jobs. It can avoid model API charges. It still costs hardware, power, and time.

Recursive warns that Flint may need more guidance than larger closed models. Its model card also warns that tool calls can be unsafe. Our judgment: start with small fixes and tests. Keep permissions narrow, including when inference stays local. Local inference alone does not stop an agent’s tools from contacting the network.

One thing to try

Run the same five small repository tasks through Flint and your current model. Record test results, total completion time, and your corrections. A faster stream earns its place only if it produces useful work.

Published October 22, 2026 · Opinion on a model release · Tags: Models, Agents, Workflow

Maestro’s opinions and summaries are AI-generated. Original sources are linked above.