Apple Silicon LLM Benchmarks

Each benchmark measures tokens per second (tok/s) during the generation phase — this is the sustained speed at which the model outputs text, excluding the time spent processing the input prompt. TTFT (time to first token) captures the initial latency before generation begins. The reference protocol for a measured row is a standardized 256-token input prompt, 512 output tokens, Q4_K_M quantization and default context settings, averaged over 3 consecutive runs on an otherwise idle machine — what vendors and contributors are asked to report. Estimated rows are computed from memory footprint and bandwidth instead; the mark on each speed cell says which you are looking at.

Each engine uses a different inference backend with distinct optimizations. MLX is Apple's native framework, purpose-built for Metal GPU acceleration on Apple Silicon — it often delivers the fastest results, especially for smaller models. Ollama uses llama.cpp with Metal support and provides reliable, consistent performance. LM Studio also wraps llama.cpp but adds a GUI layer that can introduce minor overhead. The performance gap between engines is typically 5-15% for the same model and hardware configuration.

It depends on your target model size. For small models (3-9B), even an M1 with 16 GB delivers usable speeds (40-80 tok/s). For mid-size models (14-35B), the M4 Pro with 24 GB is the sweet spot — enough RAM for 14B models at 35-55 tok/s. For large models (70B+), the M5 Max with 128 GB is ideal, offering 614 GB/s memory bandwidth. The Mac Studio M5 Ultra (1.2 TB/s) and M3 Ultra (819 GB/s), with up to 512 GB, handle the biggest models but are overkill for anything under 70B.

Yes, we welcome community submissions. Run your benchmark using Ollama, LM Studio, or MLX with standard settings (Q4_K_M quantization, default context). Record your chip model, total RAM, engine version, and both tok/s and TTFT values. Submit via our GitHub repository or by email. We verify all submissions against known performance baselines before adding them to the database.

According to the LLMCheck index as of August 2026, Maple Preview 20B-A1B is the fastest entry at 281 tokens per second on an M5 Pro (vendor-reported), with LFM2.5-2.6B at 220 tok/s on M5 Max (vendor-reported) and Gemma 4 E2B the fastest pure-estimate entry at ~261 tok/s. Among larger models, Qwen 3.6-27B (77.2% SWE-bench Verified) generates ~30 tok/s estimated on an M5 Max, and DeepSeek V4 Flash (284B-A13B MoE) achieves ~39 tok/s community-reported on a 128 GB M5 Max at 2-bit.

Memory bandwidth determines how fast your Mac can feed model weights to the GPU during inference. The LLMCheck index shows a near-linear relationship: the M5 Max (614 GB/s bandwidth) generates tokens roughly 3x faster than an M2 Pro (200 GB/s). This is why Unified Memory architecture gives Apple Silicon an advantage — there's no CPU-to-GPU transfer bottleneck.

The LLMCheck Score is a 0–100 composite metric: 50 points for model capability (sourced from Arena AI ELO, MMLU, and coding benchmarks), 25 points for Mac-specific speed (tok/s on M5 Max), 15 points for accessibility (minimum RAM), and 10 points for license openness. Full formula and per-model sources at /methodology.html.