llamaperf

Qwen3.8 27B

on M3 Max 48GB · oMLX

Tone: mixedApple hardware
Nov 7, 2026
Throughput
100.0 t/s gen
Quant
oQ4e-mtp (MLX)
System RAM
48 GB

Use cases

agentictool-usecodinglong-context

Summary

User reports Qwen3.8 27B and Qwen3.6 35B running at around 100 tk/s on an M3 Max 48GB MacBook. Setup is oMLX with MLX oQ4e-mtp models, with MTP speculation enabled. The figure is given as an approximate single value, not a measured benchmark. On a Proxmox server with an RTX 2060 Super 8GB, i9-9900K and 128GB DDR4, Qwen3.6 35B and Qwen3 Coder 30B GGUF models run at 20-40 tk/s under llama.cpp. User is asking for engine and model recommendations for agentic, long-horizon tool-calling work and is considering streaming experts from SSD on both machines.

Compare this run

Medians of plain runs: one device, one request, no speculative decoding, the whole model in the device's memory. The report above is one setup, so read it beside them and not as one of them.

Qwen3.8 27B at 4-bit by hardware

Bars are medians on one scale, 0 to 94.0 tokens/s. Each dot is one run and opens its report. Press ▶ to watch text arrive at a row's speed.

Everything reported on the M3 Max 48GB →Every Qwen3.8 report →Estimate Qwen3.8 on this card →