apple silicon
apple-silicon-llm-benchmarks
Apple Silicon Local LLM Benchmarks — M2 Max 32GB
Measurements taken while trying to get Qwen3.8-27B usable locally on a 32GB M2 Max. Most of
the popular speedup advice did not transfer from CUDA, so these are mostly negative results.
Everything here was measured on one machine. Treat it as a datapoint, not a law.
Hardware and software
Chip
Apple M2 Max
Unified memory
32 GB (~21.8 GB wireable to the GPU)
macOS
26.5.2
llama.cpp
build c1d0e7a00… See the full description on the dataset page: https://huggingface.co/datasets/RaynarDM/apple-silicon-llm-benchmarks.qwen3.8-27b-apple-silicon-concurrency
Qwen3.8-27B concurrent serving on Apple Silicon — benchmarks, patches & recipes
Research artifacts from making 16 concurrent Qwen3.8-27B requests work on a
single Apple Silicon Mac (M-series, 128 GB unified memory) — across three
serving engines, including the patches that make SGLang's native MLX
backend serve this model for the first time.
Code / full history: https://github.com/bluehawana/Qwen3.827B-SGLang-mpbm5max
Run it with Ollama: ollama run bluehawana/qwen3.8-27b-q8
Fast… See the full description on the dataset page: https://huggingface.co/datasets/bluehawana/qwen3.8-27b-apple-silicon-concurrency.humaneval-apple-silicon
Mac Coding Bench Results v1 — Speed + Code Quality Benchmarks on Apple Silicon
Speed and code quality benchmarks for quantized LLMs running locally on Apple Silicon Macs. The dataset pairs inference speed measurements (tokens/sec) with HumanEval+ functional correctness scores for 21 models, across three hardware configurations (M1, M2 Max, M5) totaling 123 benchmark results.
Key Highlights
Qwen 3.6 35B-A3B achieves 89.6% HumanEval+ pass@1 at 16.7 tok/s — best quality… See the full description on the dataset page: https://huggingface.co/datasets/enescingoz/humaneval-apple-silicon.mlx-to-gguf-apple-silicon
mlx-to-gguf-apple-silicon
Convert an MLX 4-bit fine-tuned model (the output of mlx_lm.fuse) into a
standard safetensors model that llama.cpp can convert to GGUF — so you can
run your MLX-trained model in Ollama on a Mac.
This solves a long-standing gap for the Apple Silicon local-AI community:
fine-tuning with MLX is the natural workflow on a Mac, but the resulting
quantized model could not be converted to GGUF for Ollama. The existing
solutions (e.g. Unsloth) require an NVIDIA… See the full description on the dataset page: https://huggingface.co/datasets/Car5577/mlx-to-gguf-apple-silicon.