The local coding field guide

Resources

Choose a local coding model for your Mac, connect your client, or find an answer about memory and context.

Model guides: 4
Articles: 5
Client setup guides: 6
Models covered: 4

An unfolded paper map has orange pins joined by a winding route across its folds.

Which guide do I need?

Setup guides

Connect your coding client to a local model on your Mac, then check the setup guide for its connection steps and limits.

An orange bridge bearing pad sits between a pale deck block and a pier cap.

Use Claude Code with a local LLM on Mac

Connect Claude Code to a local LLM on Mac through Bridge1. Use the generated launch command, choose Route1 and check Auto mode limits.

Tested end to end

A triangulated steel truss section has an orange diagonal member and rests on a plain base.

Use OpenCode with a local LLM on Mac

Set up OpenCode with a local LLM on Mac using Bridge1. Copy its provider entry and check prompt rereads, plan mode and stored-key behavior.

Tested end to end

Two featureless clay walkers cross a stone footbridge, with an orange backpack on the trailing traveller.

Use Pi.dev with a local LLM on Mac

Try Pi.dev with a local LLM on Mac using Bridge1. Early support is a one-turn smoke test; copy the generated custom-provider JSON and review it.

Early support · one short smoke-test turn

A matte toy train crosses slender stone viaduct arches along an orange track.

Use Codex with a local LLM on Mac

Connect Codex CLI or Desktop to a local LLM on Mac with Bridge1. Configure the Responses provider and choose Route1 or a plain model id.

Tested end to end

A featureless clay cyclist carries an orange satchel across a faceted stone bridge.

Use Hermes Agent with a local LLM on Mac

Try Hermes Agent with a local LLM on Mac through Bridge1. Early support covers a one-turn smoke test; configure the custom endpoint and key.

Early support · one short smoke-test turn

A folded paper sailboat with an orange sail passes beneath a faceted stone bridge arch.

Use Claude Desktop with a local LLM on Mac

Use Claude Desktop with a local LLM on Mac through Third-Party Inference. Set up the gateway, test the connection and review Code tab limits.

Tested end to end

Start here

Your client.
Your local model.

Select a model in Bridge1, then follow your client’s connection guide.

Model guides

Choose a local coding model for your Mac by checking its catalog profiles, memory requirements and client setup.

A small dark bird walks along an orange lane on a stone bridge across a rocky gorge, with trees at either end.

Featured guide

2026-10-20 9 min read

Ornith 1.5 35B-A3B on Mac: coding default and limits

Choose Ornith 1.5 35B-A3B when you want Bridge1’s default for coding and have memory for the selected profile. Its published Claude Code result is 40 of 40 small-repository tasks; that supports trying it on bounded coding work, not assuming it will finish every job in your own project.

Ornith 1.5 35B-A3B
MLX
GGUF / llama.cpp
Coding defaults
Memory planning

A long stone arcade has a few orange-lit arches among many quiet openings.

2026-10-19 8 min read

Qwen3.6 35B-A3B on Mac: memory and coding setup

Run Qwen3.6 35B-A3B through Bridge1 with OpenCode, Codex or Claude Code. Check memory at long context, the GGUF build and client limits.

Qwen3.6 35B-A3B
MLX
GGUF / llama.cpp
Growing chats
Memory planning

A narrow orange window frame and a wide frame stand before the same solid dark block.

2026-10-19 8 min read

Qwen3.8 27B on Mac: choose between two profiles

Choose Qwen3.8 27B MLX or GGUF by admitted context and memory evidence. Separate native context from Bridge1’s window, then connect your client.

Qwen3.8 27B
MLX
GGUF / llama.cpp
Context windows
Memory planning

A textured bonsai in a ceramic pot stands on a stone bridge deck beside an orange lane.

2026-10-19 7 min read

Bonsai 2 27B on Mac: small download, larger footprint

Run Bonsai 2 on Mac: 8.6 GB download, 53.0 GB estimated at 81,920 tokens on the reference 64 GB M1 Max. Check measured depth and client setup.

Ternary Bonsai 2 27B
MLX
Compressed weights
Memory planning

Articles

Read about memory, context and the wait before a local coding model replies on your Mac.

A clay traveller with an orange backpack starts along a winding path marked by orange stakes toward a stone footbridge, beside a folded map and brass compass.

Pi.dev and OpenCode for beginners: your first project

Ask a local model to write a page for a made-up reading club, then check the saved file yourself. Use OpenCode through Bridge1, with a practice folder for the page and a few notes to summarize.

MLX
First project
Terminal basics

A blank framed picture card passes through a brass letterbox in a stone wall with an orange ribbon trailing behind.

2026-10-29 9 min read

Local coding model vision: will it receive my screenshot?

Choose an exact profile with published image support, preserve client capability settings and distinguish an image response from a text-description response.

Start with the profile, rather than the family name
Other supported images still have coding limits

A toy cargo truck carries a stone crate, a stack of paper and a small orange crate across a stone bridge.

2026-10-29 6 min read

Why model download size is not its Mac memory requirement

Plan local model memory on Mac by separating download size, conversation growth and other apps, with measured depth kept apart from window estimates.

MLX
GGUF / llama.cpp
Memory planning
Context windows

Clay jars with blank ceramic tags and orange strings hold different amounts of material.

2026-10-29 9 min read

Read MLX and GGUF labels before choosing a model

Understand model names, MLX and GGUF labels, precision and profile-specific limits before choosing a local coding model on your Mac.

MLX
GGUF / llama.cpp
Model profiles
Quantization

Stone bars of different heights stand beside a blank-dial stopwatch and a thin ruler, with the tallest bar orange.

2026-10-29 9 min read

SWE-bench and local quants: how to read model results

Read SWE-bench scores for local model quants alongside speed evidence. Keep the tested weights, hardware, workload, repeats and missing conditions attached.