Ollama 0.40 on Apple Silicon: `mlx runner failed: panic: mlx: Maximum threads per threadgroup is 896 but requested 1024` after about a minute

Context

Source(s): https://github.com/ollama/ollama/issues/18846 Status: reported working by source, not reproduced by us (workaround only; root fix not confirmed).

Problem

With an MLX model (for example qwen3.6:35b-mlx, qwen3.6:27b-mlx, qwen3.8:27b-mlx), a longer generation or a follow-up request after a tool call crashes the MLX runner and the request returns HTTP 500. Same model and workload ran fine on Ollama 0.34.2 to 0.35.1.

Error (redacted / normalized)

mlx runner failed: panic: mlx: Maximum threads per threadgroup is 896 but requested 1024 for kernel
sdpa_vector_2pass_1_bfloat16_t_256_256_nomask_qnt_c_nosinks_128. at
.../mlx-c-src/mlx/c/transforms.cpp:15

panic: mlx: Maximum threads per threadgroup is 896 but requested 960 for kernel sdpa_vector_2pass_1_bfloat16_t_256_256_nomask_qnt_c_nosinks_128. at .../mlx-c-src/mlx/c/transforms.cpp:15

panic: cache: offset 2804 exceeds 2803 stored keys

Applies when

macOS on Apple M1 Max / M1 Pro, Ollama 0.40.0, 0.40.1-rc0 and 0.40.3, MLX-format models, large context (64k).

Confirm it is this problem

1. ollama --version → expect: 0.40.x.
2. The model tag ends in -mlx and server.log shows mlx runner failed: panic: mlx: Maximum threads per threadgroup is 896 ....

Not this when / boundary

Plain out-of-memory errors when context is too large (the reporter had to lower context to 64k to avoid OOM first). Non-MLX models.

Root cause

The MLX attention kernel in this Ollama release asks for more GPU threads per threadgroup than the M1 Pro/Max GPU allows (896) (our inference from the error text). (Maintainers did not explain further; the issue was closed on 2026-10-12 but the thread does not name a fixed release.)

Fix

1. Switch to the q4_K_M tag of the same model so Ollama uses the llama.cpp runner instead of MLX, for example ollama run qwen3.6:35b-a3b-q4_K_M (find tags at https://ollama.com/library/<model>/tags). → expect: suggested by a collaborator; one commenter confirmed it works with their production workload, another ran 34 rounds of the same conversation without an error using qwen3.6:27b-mtp-q4_K_M.
2. Or downgrade Ollama to 0.35.1, where the original poster saw no crash.
3. our note: check the release notes of Ollama versions after 0.40.3 for an MLX fix before going back to the -mlx tags.

How to verify

Run the same long prompt or tool-calling conversation → expect: it completes, no mlx runner failed in server.log.

What did not work

Deleting the Ollama config folder for a fresh setup and re-downloading the models did not help.

Risks

A commenter measured lower output speed on the llama.cpp runner (13.4 tok/s, against 23.7 tok/s with the MLX tag on 0.34.x).

Verified scope

Ollama 0.40.0 to 0.40.3 on M1 Pro/Max, October 2026.

Becomes obsolete when

An Ollama release confirms the MLX threadgroup fix.