MTPLX 2.13.1 fixes Inkling 2 Mini failing to load on M1 Macs. In 2.13.0 the model stopped while loading on an M1 Max with the error "Thread group size (800) is greater than the maximum allowed threads per threadgroup (704)" (#706). The M5 loads it. On the M4 we own, a 16 GB Mac too small for any Inkling 2 Mini pack, the GPU runs every Inkling 2 Mini kernel at the model's sizes, but no full load was run there. Whether Macs with an M2 or an M3 were affected is not known: nobody has run 2.13.0 on one yet. The packs on Hugging Face did not change, so nothing needs to be downloaded again.
| What you do | 2.13.0 | 2.13.1 |
|---|---|---|
| Load Inkling 2 Mini on an M1 Max (Optimized Quality FP16, as in #706) | Fails with "Thread group size (800) is greater than the maximum allowed threads per threadgroup (704)" | Loads |
| Run Inkling 2 Mini on a 16 GB M4 | Every Inkling 2 Mini kernel and the Inkling test files ran at the model's sizes; no full load | Every Inkling 2 Mini kernel and the Inkling test files ran at the model's sizes; no full load |
| Load Inkling 2 Mini on an M5 | Loads | Loads. The three published packs give the same tokens as 2.13.0, and an interleaved pair is flat: decode within 2%, first token equal or lower, prefill and memory unchanged |
How the M1 row was checked: on GitHub's M1 test machine, whose GPU has the same limit of 704 threads for this kernel, 2.13.0's code is refused in 14 of 106 wide launches and 2.13.1 runs all 106. On an M5 Max set to the M1 Max's limit, all six packs load and answer, and Optimized Quality FP16 gives the same tokens as 2.13.0 in ten cases. It has not yet been run on a real M1 Max. If Inkling 2 Mini still does not load on your M1, M2 or M3 Mac, please open an issue with the output.
What changed
- Inkling 2 Mini now loads on M1 Macs. While it loads, Inkling 2 Mini prepares the widest panel it uses when the model copies text it has already seen, such as rewriting a file: 25 rows, 24 copied tokens and one more. That panel's attention kernel asks the GPU for a group of 32 threads per row, 800 threads, and an M1 Max allows that kernel only 704. The first time each large kernel runs, MTPLX now checks that this Mac's GPU accepts its thread group. Where the GPU refuses the wide panel, its attention runs in groups of 8 rows, which gives exactly the same result, so answers and the 24-token copy rounds are unchanged. Any other kernel a GPU refuses falls back to the standard path for that step instead of stopping the load, and
mtplx doctor --explainand /health list it. - A source install builds MTPLX's engine with older Xcode tools. In 2.13.0 a source install (
pipwithout a prebuilt wheel, or a development checkout) whose Xcode tools carried an SDK older than macOS 26 could not compile the engine, and MTPLX then ran without it. It now compiles there too. The prebuilt wheels on PyPI for Python 3.11 to 3.14 on Apple silicon carry the engine, and Homebrew installs that wheel.
If you have an M1 Mac
Update MTPLX, then start Inkling 2 Mini again. The app offers the update itself; with Homebrew run brew update && brew upgrade mtplx, and with pip run pip install -U mtplx. If you started 2.13.0 with MTPLX_CONTEXT_COPY_K=7 to get past the error, you can leave that setting out now. You get the same answers as on any other Mac, and one line at load says the wide panel runs in row groups. Speed on an M1 is not measured: we do not have an M1 with 24 GB or more. The packs you already downloaded stay as they are; on M1 and M2 Macs MTPLX offers the FP16 builds of the three packs, which are made for those chips.
Validated
| Check | Result |
|---|---|
| Every wide Inkling 2 Mini launch on GitHub's M1 test machine, on the release pull request | Run 45011806448: all 106 wide launches run. 2.13.0's code on the same machine: 14 of 106 refused |
The real M4 gate, on a 16 GB M4, on the fix's head f7ebc03ff (the same code as 2.13.1's fix), 22 November 2026 |
1,379 passed, 1 known failure in a test that expects an M5 |
| GitHub's checks on the release pull request | All green: ci 45011717653, build 45011717658, hygiene 45011717655, no-ai-attribution 45011717663, and the kernel matrix 45011806448 (kernel exactness, a small model loading and answering, and the Inkling launch check above) |
| Every test file, on the release pull request's code | 14,332 passed, 1 failed, 150 skipped. The one failure, a check on a short prompt asked twice, already failed before 2.13.0 was released, and this release does not touch it |
Upgrading
The app updates itself (MTPLX menu, Check for Updates). With Homebrew: brew update && brew upgrade mtplx. With pip: pip install -U mtplx.
Credits
- The report, with the whole traceback: #706 by @haseebeqx.
b65d0850e4a68bf4f95d8d1d2a5efd85966860e47179dd6884461ab4814340cb. Every published speed
number with its conditions is on the benchmarks page; the archive of every version is
on the releases page.