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 GBspeed41.241.817.4570.91
Qwen3-Coder-Next 80B-A3B (6-bit)16 GBexact6.446.3517.41291.00
Qwen3-Coder-Next 80B-A3B (6-bit)32 GBspeed34.935.015.9471.00
Qwen3-Coder-Next 80B-A3B (6-bit)32 GBexact13.914.816.1771.00
Qwen3-Coder-Next 80B-A3B (6-bit)48 GBspeed34.534.515.8441.00
Qwen3-Coder-Next 80B-A3B (6-bit)48 GBexact17.718.316.2571.00
Qwen3.8-Flash-Next 229B (6-bit)16 GBspeed17.116.937.31280.30
Qwen3.8-Flash-Next 229B (6-bit)16 GBexact3.533.5738.23731.00
Qwen3.8-Flash-Next 229B (6-bit)32 GBspeed18.318.131.51161.00
Qwen3.8-Flash-Next 229B (6-bit)32 GBexact6.186.0331.72331.00
Qwen3.8-Flash-Next 229B (6-bit)48 GBspeed19.219.432.21121.00
Qwen3.8-Flash-Next 229B (6-bit)48 GBexact7.887.7232.01861.00
Qwen3-235B-A22B (4-bit)16 GBspeed3.46-45.95630.19
Qwen3-235B-A22B (4-bit)16 GBexact1.131.1246.66051.00
Qwen3-235B-A22B (4-bit)32 GBspeed9.719.7144.61830.88
Qwen3-235B-A22B (4-bit)32 GBexact1.821.7046.64641.00
Qwen3-235B-A22B (4-bit)48 GBspeed9.199.2842.11490.98
Qwen3-235B-A22B (4-bit)48 GBexact2.622.3847.63431.00
MiniMax-M3 426B (4-bit)32 GBspeed3.533.3843.73660.38
MiniMax-M3 426B (4-bit)32 GBexact1.411.3944.78081.00
MiniMax-M3 426B (4-bit)48 GBspeed4.624.4744.92950.75
MiniMax-M3 426B (4-bit)48 GBexact2.202.1843.15261.00

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