Local LLM benchmarks on Mac
If you care about tok/s first, start here. On Apple silicon the Global Agent harness clocks an 80B-class coding MoE at 34.5 median decode tok/s on 48 GB speed settings and 41.2 even in a 16 GB sandbox, and a 229B MoE at 19.2 at 48 GB with top-1 agree 1.00. That is a working local coding loop, not a slideshow.
What developers should watch
Decode tok/s is the headline (median across requests). Tool-call tok/s tells you how the agent feels when it writes files and runs commands: 35.0 on the 80B coding MoE at 32 GB, 19.4 on the 229B at 48 GB, and 9.28 on a 235B. Exact mode trades some rate for locked fidelity (top-1 agree 1.00 on every exact run).
Measured results
| Model | RAM | Mode | Decode tok/s | Tool call tok/s | First turn s | Task s | Top-1 agree |
|---|---|---|---|---|---|---|---|
| Qwen3-Coder-Next 80B-A3B (6-bit) | 16 GB | speed | 41.2 | 41.8 | 17.4 | 57 | 0.91 |
| Qwen3-Coder-Next 80B-A3B (6-bit) | 16 GB | exact | 6.44 | 6.35 | 17.4 | 129 | 1.00 |
| Qwen3-Coder-Next 80B-A3B (6-bit) | 32 GB | speed | 34.9 | 35.0 | 15.9 | 47 | 1.00 |
| Qwen3-Coder-Next 80B-A3B (6-bit) | 32 GB | exact | 13.9 | 14.8 | 16.1 | 77 | 1.00 |
| Qwen3-Coder-Next 80B-A3B (6-bit) | 48 GB | speed | 34.5 | 34.5 | 15.8 | 44 | 1.00 |
| Qwen3-Coder-Next 80B-A3B (6-bit) | 48 GB | exact | 17.7 | 18.3 | 16.2 | 57 | 1.00 |
| Qwen3.8-Flash-Next 229B (6-bit) | 16 GB | speed | 17.1 | 16.9 | 37.3 | 128 | 0.30 |
| Qwen3.8-Flash-Next 229B (6-bit) | 16 GB | exact | 3.53 | 3.57 | 38.2 | 373 | 1.00 |
| Qwen3.8-Flash-Next 229B (6-bit) | 32 GB | speed | 18.3 | 18.1 | 31.5 | 116 | 1.00 |
| Qwen3.8-Flash-Next 229B (6-bit) | 32 GB | exact | 6.18 | 6.03 | 31.7 | 233 | 1.00 |
| Qwen3.8-Flash-Next 229B (6-bit) | 48 GB | speed | 19.2 | 19.4 | 32.2 | 112 | 1.00 |
| Qwen3.8-Flash-Next 229B (6-bit) | 48 GB | exact | 7.88 | 7.72 | 32.0 | 186 | 1.00 |
| Qwen3-235B-A22B (4-bit) | 16 GB | speed | 3.46 | - | 45.9 | 563 | 0.19 |
| Qwen3-235B-A22B (4-bit) | 16 GB | exact | 1.13 | 1.12 | 46.6 | 605 | 1.00 |
| Qwen3-235B-A22B (4-bit) | 32 GB | speed | 9.71 | 9.71 | 44.6 | 183 | 0.88 |
| Qwen3-235B-A22B (4-bit) | 32 GB | exact | 1.82 | 1.70 | 46.6 | 464 | 1.00 |
| Qwen3-235B-A22B (4-bit) | 48 GB | speed | 9.19 | 9.28 | 42.1 | 149 | 0.98 |
| Qwen3-235B-A22B (4-bit) | 48 GB | exact | 2.62 | 2.38 | 47.6 | 343 | 1.00 |
| MiniMax-M3 426B (4-bit) | 32 GB | speed | 3.53 | 3.38 | 43.7 | 366 | 0.38 |
| MiniMax-M3 426B (4-bit) | 32 GB | exact | 1.41 | 1.39 | 44.7 | 808 | 1.00 |
| MiniMax-M3 426B (4-bit) | 48 GB | speed | 4.62 | 4.47 | 44.9 | 295 | 0.75 |
| MiniMax-M3 426B (4-bit) | 48 GB | exact | 2.20 | 2.18 | 43.1 | 526 | 1.00 |
Source: Global Agent 1.2.0 bench harness on Apple M5 Pro, 2026-10-09. Each model builds the same Express ping/pong app with a web page, installs it and starts it, using the system prompt and tools the app sends. Decode tok/s is the median across the task's requests. Speed is the speed slider at 20; exact keeps every expert the router picks. 16 and 32 GB are sandboxed on the same Mac. MiniMax-M3 needs 32 GB or more. Full interactive tables: globalagent.dev/benchmarks.
Download free for Mac - Read the docs - Licenses
Questions: info@globalagent.dev
© QWERTY Code LLC - Terms - EULA - Get started - Docs