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

Reference Mac: Apple M1 Max · 32-core GPU · 64 GB. Decode is engine-reported tokens per second; other Macs are not measured.

Compare profiles in the model catalog

Select Qwen3.6 35B-A3B, MLX mxfp4, in Bridge1. Check memory before loading it. The catalog marks this profile as an Editor’s Choice; it is not the default coding model.

The 35B name counts all parameters. Qwen reports about 3B active parameters per token, but that smaller active count does not make this a 3B-sized download or memory footprint.

Swipe to see the whole row

The memory pill gives approximate start memory and growth; the context label shows an automatic limit of 229,376 out of 262,144 tokens.

Measured to 131,443 tokens (39.1 GB peak, M1 Max 64 GB). The 229,376-token window is estimated at 55.2 GB.

The catalog is dated 2026-10-09. The reference Apple M1 Max has a 32-core GPU. Your Mac’s memory budget can admit a different window.

Model weights stay in memory while attention state grows with the conversation. Leave room for macOS and your coding client as well. Do not subtract the 19.3 GB download from 55.2 GB and call the difference a measured cache size: those numbers describe different things.

Figure 1 shows the app’s approximate memory labels: about 21 GB to start, plus 1.4 GB per 10K tokens. These approximations are not measurements. Figure 2 labels the 55.2 GB figure at the default window as estimated; it does not predict a safe chat length for a smaller Mac.

Choose a smaller catalog model when keeping your editor and other apps open leaves too little memory for this profile’s admitted window. On a 32 GB Mac, the 64 GB reference result is a reason to reduce the window or choose less demanding weights, not a promise that the same long chat will fit.

On macOS, bridge1 launch starts Claude Code, Codex, OpenCode or Pi with Bridge1, subject to the installed client version, model availability and loading settings; the manual steps still work. Open the Gateway panel after selecting the model. Copy the setup for your client, including the generated model name and context settings. Route1 follows the model selected in Bridge1; use the native name supplied there when you want to target one model directly.

For Claude Code, the base URL is http://127.0.0.1:54321, without a /v1 suffix. The command below comes from the Claude Code setup guide and uses a placeholder key. Copy Gateway’s full command for your selected model, including its extra context setting for windows of 100,000 tokens or more. OpenCode and Codex use a base URL ending in /v1; follow their guides below.

ANTHROPIC_BASE_URL="http://127.0.0.1:54321" \
ANTHROPIC_API_KEY="YOUR-GATEWAY-KEY" \
ANTHROPIC_MODEL="claude-route1" \
API_TIMEOUT_MS="1800000" \
CLAUDE_STREAM_IDLE_TIMEOUT_MS="2400000" \
CLAUDE_CODE_DISABLE_NONSTREAMING_FALLBACK="1" \
ANTHROPIC_CUSTOM_HEADERS="X-Bridge1-Client: claude-code" \
claude

The GGUF UD-Q3_K_M build runs through llama.cpp. It is a separate choice in Bridge1: 17.1 GB on disk, a 262,144-token admitted window and 27.0 GB estimated memory on the reference Mac.

Consider this build when its smaller catalog memory estimate fits your budget better. Keep the estimated label in view. It is not a measured 27 GB requirement, and the build’s name alone does not establish a speed advantage or equal coding quality between the two quantizations.

Swipe to see the whole row

The memory pill marks 27.0 GB at 262,144 tokens as estimated.

As of 26 September 2026, Bridge1’s published task evaluation recorded 38 passes out of 40 for Qwen3.6 35B-A3B through Claude Code. Those were small Python repositories, with hidden tests; the result does not establish the same behavior in a large project or every client.

The catalog’s decode figures measure output generation. They do not include the whole wait to read a long prompt or finish a coding task. Use the Speed page for that distinction and the Evidence page for the test conditions; the key facts above preserve missing conditions rather than turning token rates into a response-time promise.

Qwen’s original model page declares Apache-2.0. The MLX Community mxfp4 page and Unsloth GGUF page each declare Apache-2.0 too. The three linked references identify those artifacts separately, accessed 2026-10-19.

Check client support and known limits