Builds / Apple M6 (12-core GPU, unified)
Mac mini M6 32 GB: Qwen3.6 35B-A3B at 63 tok/s, Nemotron 3.5 Lightning 45 (LM Studio)
Source: github.com ↗📏 How measured: reported by the original source; RigBase re-checked the numbers at that source
- GPU(s)
- Apple M6 (12-core GPU, unified)
- VRAM / unified memory (GB)
- 32
- CPU
- Apple M6 (12 CPU cores)
- Motherboard
- not stated
- RAM
- 32 GB unified
- PSU
- not stated
- Other
- Mac mini (Mac18,5), macOS 27.0.1; LM Studio; 32,768 context, reasoning off, max_tokens 600, temperature 0.7
- Reported total cost
- not stated
- Power
- not stated
Measured speed (generation tokens/sec)
| Model | Quant | Engine | tok/s | Prompt processing (tok/s) | Details |
|---|---|---|---|---|---|
| Qwen3.6 35B-A3B | MLX 4-bit | LM Studio (MLX) | 63.2 | not stated | 📏 Median of 3 runs per prompt; lowest of the three prose prompts (63.2-63.8) |
| Qwen3.6 35B-A3B | MLX affine 4-bit (group 64) + MTP head and DFlash2 drafter | Kapitsa 0.2.4 (KFlash tree speculative decoding) | 69.3 | not stated | ⚡ spec-decode (KFlash)📏 Median of 3 runs; lowest prose prompt (69.3-84.1; code 160.1, JSON 219.8) |
| Qwen3.6 35B-A3B | Splash package (incoai/Qwen3.6-35B-A3B-Splash) | LM Studio Splash backend | 51.5 | not stated | ⚡ spec-decode (DFlash)📏 Median of 3 runs; lowest prose prompt (51.5-69.6; code 135.9, JSON 234.1) |
| Gemma 4 26B-A4B | QAT (google/gemma-4-26b-a4b-qat) | LM Studio | 45.5 | not stated | 📏 Median of 3 runs; lowest prose prompt (45.5-48.2) |
| Nemotron 3.5 Lightning 30B-A3B | GGUF Q4_K_M (lmstudio-community) | LM Studio (llama.cpp) | 45.1 | not stated | 📏 Median of 3 runs; lowest prose prompt (45.1-45.6) |
| Ornith 1.5 35B-A3B | GGUF Q4_K_M | LM Studio (llama.cpp) | 37.6 | not stated | 📏 Median of 3 runs; lowest prose prompt (37.6-43.7) |
| Qwen3.5 9B | MLX | LM Studio (MLX) | 30.3 | not stated | 📏 Median of 3 runs; lowest prose prompt (30.3-30.8) |
| Gemma 4 12B | QAT (google/gemma-4-12b-qat) | LM Studio | 18.2 | not stated | 📏 Median of 3 runs; lowest prose prompt (18.2-18.8) |
| Qwen3.8 27B | Splash package (incoai/Qwen3.8-27B-Splash) | LM Studio Splash backend | 15.3 | not stated | ⚡ spec-decode (DFlash)📏 Median of 3 runs; lowest prose prompt (15.3-20.2; code 45.2, JSON 67.7) |
| Qwen3.8 27B | GGUF Q4_K_M | LM Studio (llama.cpp) | 11.2 | not stated | 📏 Median of 3 runs; lowest prose prompt (11.2-11.8) |
| Gemma 4 31B | QAT (google/gemma-4-31b-qat) | LM Studio | 7.4 | not stated | 📏 Median of 3 runs; lowest prose prompt (7.4-7.5) |
Measured 3-4 Oct 2026 (LM Studio table) and 22 Oct 2026 (Kapitsa) by the repo owner; raw JSONL data and script in the repo. Values shown are the lowest of the prose prompts; code and JSON prompts run much faster with speculative decoding.
Source: https://github.com/HDZucht/Mac-Mini-M6-LLM-benchmarks
Blank fields = not stated in the source. We never estimate build data; GPU street prices above are separately sourced and dated.
Rent before you buy: try this model on a cloud GPU for a few dollars first: Runpod (referral link; sign up with Google to get the new-user bonus)