Can a 32 GB Mac Run a Local Coding Agent? Codex and Cline, Measured
We gave Codex CLI and Cline two real Python tasks each on a 32 GB Mac mini, with a 27B model running locally. All four passed. Here is how long they took, where the time went and what to set up.
Yes. On a 32 GB Mac mini (M2 Pro), Codex CLI and Cline completed 4 of 4 real coding tasks with qwen3.8-27b-4bit running locally through Rapid-MLX. A task took 2.5 to 6.5 minutes, and peak GPU memory stayed at 23.5 GB. The first request on a fresh server is the slow part, about one to one and a half minutes. After that, the median step took 18 seconds.
The result
Two tasks on the same small Python package (an inventory store with bugs), run once per agent in a fresh copy of the repo:
- Task A, fix failing tests: three bugs in
inventory/store.py, three failing tests. The agent must fix the code, not the tests. - Task B, add features with tests: add
Store.restock()with error handling, a newinventory/export.pywithto_csv(), a publicStore.items(), and new tests; keep the old tests passing. We then ran 4 hidden tests the agent never saw.
| Agent | Task | Result | Wall time | Model requests | Peak GPU memory |
|---|---|---|---|---|---|
| Codex CLI 0.169.1 | A · fix tests | 3/3 tests pass | 160 s | 4 | 23.4 GB |
| Cline CLI 3.0.74 | A · fix tests | 3/3 tests pass | 246 s | 8 | 19.3 GB |
| Codex CLI 0.169.1 | B · add features | 11/11 own tests, 4/4 hidden | 394 s | 14 | 23.5 GB |
| Cline CLI 3.0.74 | B · add features | 8/8 own tests, 4/4 hidden | 259 s | 8 | 23.5 GB |
Setup: Mac mini, M2 Pro, 32 GB, macOS. rapid-mlx 1.0.4, model qwen3.8-27b-4bit (the engine's pick for 32 GB), default settings, no flags. One run per agent and task, measured on 2026-11-17. Request counts and memory come from the server log; GPU memory is the Metal peak the server reports.
Where the time goes
A coding agent sends a long system prompt with its tool definitions before your task even starts: 3,400 tokens for Cline and 5,400 for Codex in these runs. The model has to read all of it before it writes the first word.
| What happens | Measured on M2 Pro 32 GB |
|---|---|
| Time to first token, first request on a fresh server | 56 s for 3,406 tokens (Cline), 90 s for 5,436 tokens (Codex): about 60 prompt tokens/s |
| Whole first request | 64 s (Cline), 96 s (Codex) |
| Later requests, prompt already cached (31 requests) | median 18 s; half took 14–38 s; range 8–73 s |
| Writing output (decode), cached requests | about 9–10 tokens/s |
The later requests are fast because Rapid-MLX keeps the prompt it has already read in its prefix cache. Each new request only processes the part that changed. In Codex task A, the second request was 6,042 tokens long, but only 634 of them were new. The longest follow-ups (52–73 s) were the steps where the agent wrote 530–730 tokens of new code. Task B's first Codex request took only 31 s because it ran on the same server right after task A, so the system prompt was already cached.
Keep rapid-mlx serve running between tasks. Restarting the server empties the prefix cache, and the next request pays the full prompt-reading cost again.
1 · Install Rapid-MLX and an agent
curl -fsSL https://rapidmlx.com/install.sh | bash npm install -g @openai/codex # or: npm install -g cline
2 · Serve the 27B model
rapid-mlx serve qwen3.8-27b-4bit
The first run downloads the weights. On a 32 GB Mac, close other memory-heavy apps; the model plus an agent session peaked at 23.5 GB of GPU memory in our runs.
3 · Point the agent at it
For Codex CLI:
rapid-mlx agents codex --setup
For Cline CLI:
cline auth -p openai -b http://localhost:8000/v1 -k not-needed -m qwen3.8-27b-4bit
In the Cline VS Code extension, choose OpenAI Compatible and enter the same base URL, any key and the model name. Full details are in the Codex guide and the Cline guide.
4 · Give it a task
# inside your project folder codex "The tests in tests/test_store.py are failing. Fix the code, not the tests. Verify with python -m pytest -q"
Ask for a check the agent can run, like a test command. In our runs, both agents read the code, edited it, ran pytest and only finished once the tests passed.
What to expect day to day
- It works, at a walking pace. A small bug fix took 2.5–4 minutes end to end. Start a task and check back, rather than watching it.
- Bigger tasks need more requests. Task B took up to 14 requests. Each one after the first reused the cached prompt, and steps that wrote more code took longer.
- Memory is the limit on 32 GB. A 27B model fits with room for an agent session. A 35B model needs 48 GB.
- Your code stays on your Mac. Requests go to
localhost.
Frequently asked questions
Is 32 GB enough RAM for a local coding agent on a Mac?
Yes, for a 27B model. In our runs, Codex CLI and Cline on qwen3.8-27b-4bit passed all four tasks on a 32 GB M2 Pro, peaking at 23.5 GB of GPU memory. With 16 GB, use a smaller model such as qwen3.5-9b-4bit; we haven't run these two tasks on it.
Why is the first request so slow?
The agent sends 3,400–5,400 tokens of instructions and tool definitions first. On an M2 Pro, the first token arrived after 56–90 seconds, about 60 prompt tokens per second for a 27B model. Later requests reuse the cached prompt: the median took 18 seconds.
Which is faster locally, Codex CLI or Cline?
It depends on the task. Cline sends a shorter system prompt, so its first request was faster (64 s vs 96 s). On the bug-fix task Codex finished sooner overall (160 s vs 246 s), and on the feature task Cline did (259 s vs 394 s). Both passed everything.
Will a newer Mac be faster?
Prompt reading scales with GPU compute, so M4 Pro, M5 and Max chips cut the first-request wait. See local AI speed across M2 Pro, M4 Pro, M5 Max and M3 Ultra for measured numbers.
Where to go next
- Best local LLM for your Mac by RAM: the pick for each memory size.
- Local coding agents on Apple Silicon: every agent Rapid-MLX sets up.
- All agent guides: Claude Code, OpenCode, Aider, Goose and more.
rapid-mlx serve <alias> serves any model on
localhost:8000/v1.
curl -fsSL https://rapidmlx.com/install.sh | bash
New models and speedups, in your inbox
A short note whenever Rapid-MLX gets faster or adds models worth running on your Mac. No spam — unsubscribe anytime.