community pool / model answer

Qwen3.6-27B-MTP-GGUF

chat · unsloth/Qwen3.6-27B-MTP-GGUF

66.4 tok/s · model median · 161 runs in the model record.

Reference rig	Quantization	Median result	Basis	TTFT (ms)	Peak VRAM	Max context	Engines
Arc Pro B70 32GB	4-bit	30.7 tok/s	reported	500	- GB	4,096	llama.cpp
Arc Pro B70 32GB	8-bit	36.1 tok/s	reported	785	- GB	32,768	llama.cpp
M5 Pro 24GB	2-bit	21.5 tok/s	reported	150	- GB	2,048	llama.cpp
M5 Pro 24GB	3-bit	20 tok/s	reported	164	- GB	2,048	llama.cpp
M5 Pro 24GB	4-bit	18.2 tok/s	measured	175	- GB	2,048	llama.cpp
Radeon AI Pro R9700 32GB	4-bit	43.3 tok/s	reported	77	- GB	2,048	llama.cpp
RTX 3060 12GB ×2	4-bit	17.7 tok/s	measured	7,257	18.1 GB	131,072	llama.cpp
RTX 3090 24GB	1-bit	49.2 tok/s	measured	-	2.5 GB	196,608	llama.cpp
RTX 3090 24GB	4-bit	101.5 tok/s	measured	761	20.2 GB	131,072	llama.cpp
RTX 3090 24GB ×2	4-bit	65.2 tok/s	measured	497	- GB	262,144	llama.cpp
RTX 4090 24GB	4-bit	94.1 tok/s	reported	-	23.9 GB	76,000	llama.cpp
RTX 5060 Ti 16GB	3-bit	26.7 tok/s	reported	596	- GB	2,048	llama.cpp
RTX 5060 Ti 16GB ×2	4-bit	33.3 tok/s	reported	179	19.3 GB	65,536	llama.cpp
RTX 5090 32GB	5-bit	70.4 tok/s	reported	141	- GB	2,048	llama.cpp
RTX PRO 6000 Blackwell 96GB	6-bit	60.1 tok/s	reported	301	- GB	131,072	llama.cpp
RTX PRO 6000 Blackwell 96GB	16-bit	40.3 tok/s	reported	282	- GB	131,072	llama.cpp
RX 7900 XTX 24GB	4-bit	48.5 tok/s	measured	1,359	21.9 GB	262,144	llama.cpp
Ryzen AI Max 395 128GB	4-bit	17.1 tok/s	measured	29,410	21.2 GB	65,536	llama.cpp
Ryzen AI Max 395 64GB	4-bit	29.2 tok/s	measured	4,063	18.3 GB	8,192	llama.cpp

Context tested

The table above is the tested context for each exact rig and quantization cell. Empty fields remain empty rather than being filled with an estimate.