vllm-mlx

High-performance OpenAI and Anthropic compatible LLM inference server for Apple Silicon. Native MLX, continuous batching, multimodal models, MCP tool calling, and Claude Code support.

Runs with mockssource: GitHubhomepagePythonApache-2.0commit 37a16c76bb21

Python, Apache-2.0 licensed. The project labels itself: anthropic, anthropic api, apple silicon, claude code, continuous batching, inference server, llm and local llm.

vllm-mlx runs, with stand-ins for the services it depends on. An Argusic agent installed it in 30 minutes and hit 2 errors and fixed every one without help on a clean machine with no GPU, and the whole session was recorded.

vllm-mlx 0.5.0 installed in a venv with all dependencies; 1585 Linux-compatible unit tests pass, 19 skip on non-Apple-Silicon, 0 failed; all 4 CLI entry points print help successfully.

What the agent ended up with on a clean machine, in its own words. How this is measured

Measured by Argusic on a fresh machine every time. Every number links to its evidence. Argusic Score 92 of 100 (the mean of this project's run scores).

At a glance

verdict
Runs with mocks
Argusic Score
92 / 100
cost of the verifying run
$0.11 (measured)
recorded runs
2
last tested
stars
1,656
forks
238
open issues
107
watchers
11
size
16 MB
created
last push

Subject data from GitHub, linked at the top of this page, refreshed . Test data by Argusic (CC BY 4.0); every number links to a run page with the full log, the recording, and their sha256 hashes.

What was measured

From the run that decided this verdict. Every line is a recorded value, not an estimate.

Time lost to each failure while testing vllm-mlx, drawn from the recorded run
Every bar is a failure the agent recorded, with the time it cost. Drawn from this run, not from anywhere else.

What went wrong, in the agent's own words

The agent's own words, worst first, never rewritten into advice.

  • 4 test files (test_paged_cache_benefits, test_paged_cache_real_inference, test_paged_cache_real_model, test_registry_idle_unload_real_model) called sys.exit(0) at module level on non-Apple-Silicon, crashing pytest collection8 minutes
  • mlx native library (libmlx.so) cannot load on x86_64 Linux, causing ImportError on all Apple-Silicon-only test files that import mlx.core at module level2 minutes
Install time
30 minutes
Cold machine to finish
19 minutes
Errors hit and fixed
2 hit, 2 fixed with no human help
How the result was proved
pytest on CI's Linux whitelist (42 files, -k 'not Integration and not InjectJson and not TestMLXMultimodalLMCache'): 1585 passed, 19 skipped, 0 failed. MLX-stubbed scheduler test: 6 passed, 0 failed
Model tokens used
103,312
Exact commit tested
37a16c76bb21
Container image
sha256:cdd920bce783

What was not measured

The other half, so you can see exactly how far this evidence goes. A gap is printed as a gap and never filled in.

Peak GPU memory
this ran on a CPU machine, so there was no GPU to sample
Behaviour over a long run
each attempt uses the smallest documented settings and a fixed time budget, so this measures whether it runs, not how it holds up under load
Behaviour against real services
this run was verified against stand-in services only

Watch the session

The whole run, from a clean machine, at the speed it actually happened. Nothing is cut and nothing is sped up.

A recording of attempt 1, run . The verdict above comes from a different attempt, so read this as one session and not as the current result. The terminal recording it was made from is the evidence, stored with a sha256 fingerprint; the video is a rendering of it. Nothing loads from the video host until you press play.

The end of the recording

The last lines of the terminal session that decided this verdict, exactly as they were recorded.

1. **`sys.exit(0)` in 4 test files crashed pytest collection** — `test_paged_cache_benefits.py`, `t…
2. **`libmlx.so: cannot open shared object file` on 11 test files** — `pip install -e ".[dev]"` pul…
**Verification:**
- **`vllm-mlx --help`** prints usage with all subcommands (serve, bench, download, model, bench-ser…
- **`vllm-mlx serve --help`** prints the full 311-line options reference
- **`vllm-mlx-chat --help`**, **`vllm-mlx-text-chat --help`**, **`vllm-mlx-bench --help`** all work
- **Linux test suite**: 1585 passed, 19 skipped, 0 failed (matching CI's whitelist)
- **MLX-stubbed scheduler test**: 6 passed, 0 failed
The server cannot actually start on this x86_64 Linux machine because it requires Apple Silicon + M…
2026-11-14T01:42:50.938467Z ERROR codex_core::tools::router: error=cannot update goal because this …
tokens used
103,312

Replay the whole session, every command from a clean machine to this point.

How it was tested

One agent, one container, no human help. It installs the project the way its own documentation says to, runs it, and fixes what breaks. Everything below is recorded as it happened: the terminal session, the log and the exact commit. The full method.

Strengths and limits

Measured facts, not opinions. How this is written.

What went well

  • Recovered from all 2 errors without a human stepping in, which says the failures are documented well enough to solve.
  • Licensed Apache-2.0, as reported by its host.
  • Measured 2 times, so the result is not a one-off.

What did not, or is not known

  • Only reached a running state against stand-in services, so nothing here proves it works against the real ones.
  • Took 30 minutes to install, slower than the median of the 9 comparable projects Argusic has measured.

Of the 10 anthropic projects Argusic has installed and timed, vllm-mlx was the 8th fastest to reach a running state, and 5 of 10 reached one at all.

Also tested, in the same area

Every one of these was installed and run by Argusic on a clean machine. Nothing appears here that was not tested.

Run history

tested byattemptoutcomescorecost USDwhen
Argusic Runner2/3Runs with mocks92.000.11
Argusic Runner1/3Runs with mocks92.000.10

Topics (from GitHub)

anthropicanthropic-apiapple-siliconclaude-codecontinuous-batchinginference-serverllmlocal-llmmacosmcpmlxmultimodal-aiopenaiopenai-apiopenai-compatiblespeech-to-texttext-to-speechtool-callingvision-language-modelvllm

Embed the badge

Markdown for the project README. It links back here; terms on the terms page.

[![Tested by Argusic](https://argusic.com/badge/vllm-mlx.svg)](https://argusic.com/subject/vllm-mlx)

Questions

Does vllm-mlx run?
vllm-mlx runs, with mocks standing in for external services. Argusic installed and launched it on a clean machine in 30 minutes, hitting 2 errors on the way, and recorded the session.
How did Argusic test vllm-mlx?
On a fresh, disposable machine, with every command recorded and the repository pinned at commit 37a16c76bb21. 2 attempts are recorded, and the full method is on the methodology page.
What did it cost to test vllm-mlx?
The run that produced this verdict cost $0.11: real compute and model cost, not a list price per million tokens. It is recorded on that run's page.
How long does vllm-mlx take to install?
30 minutes in the run that decided this verdict, from a clean machine to a running state. That is one measured attempt, not an average.
Does vllm-mlx need a GPU?
Not to start. Argusic reached a running state on a machine with no GPU.
What goes wrong when installing vllm-mlx?
2 things broke in the recorded run, and 2 were fixed without human help. Each one, and the time it cost, is listed on this page.
How does vllm-mlx compare with the alternatives?
Of the 10 anthropic projects Argusic has installed and timed, vllm-mlx was the 8th fastest to reach a running state, and 5 of 10 reached one at all.
Where is the evidence for vllm-mlx?
All 2 recorded runs are on this page, each linking to its full log and terminal recording, stored with a sha256 fingerprint so it cannot be quietly altered.

Discussion