Configs › Qwen3.8-Flash-Next tok/s: Mac Ultra 76.4, Mac Max 57.5

Qwen3.8-Flash-Next tok/s: Mac Ultra 76.4, Mac Max 57.5

Measured Qwen3.8-Flash-Next speeds: 76.4 tok/s on Mac Ultra; 57.5 tok/s on Mac Max (Q4-INF, mlx). 22 configs, each linked to its source, snapshot 2026-10-16.

Snapshot 2026-10-16 · 394 configs on AI boxes and 423 on GPU cards · 123 models. Every number is generated from the tracker snapshot and linked to where it was measured; nothing here is typed by hand.

Quick answers

How fast is Qwen3.8-Flash-Next on Mac Ultra?: 76.4 tok/s single-stream decode (MLX-4BIT, mlx (speculative decode, mtp)), the only measured config on Apple Mac Studio (M-series Ultra), measured by scouzi1966, source last checked 2026-09-06. source
How fast is Qwen3.8-Flash-Next on Mac Max?: 57.5 tok/s single-stream decode (Q4-INF, mlx), the fastest of 2 measured configs on Apple M-series Max (MacBook Pro / Mac Studio Max), measured by inferencerlabs, source last checked 2026-09-06. source
How fast is Qwen3.8-Flash-Next on Strix Halo?: 55.3 tok/s single-stream decode (llama.cpp/hip (speculative decode, mtp)), the fastest of 11 measured configs on AMD Strix Halo (Ryzen AI Max+ 395), measured by abliter8-ai, source last checked 2026-09-18. source
How fast is Qwen3.8-Flash-Next on DGX Spark?: 44.2 tok/s single-stream decode (NVFP4, vllm (speculative decode, mtp-2)), the fastest of 8 measured configs on NVIDIA DGX Spark (GB10), measured by YSLAB-ai, source last checked 2026-09-01. source

Measured configs for Qwen3.8-Flash-Next on AI boxes

Model	Hardware	Quant	Backend	Decode	Context	Trust	Source
Qwen3.8-Flash-Next	Mac Ultra	MLX-4BIT	mlx (speculative decode, mtp)	76.4 tok/s		Extracted, source-linked	source
Qwen3.8-Flash-Next	Mac Max	Q4-INF	mlx	57.5 tok/s		Extracted, source-linked	source
Qwen3.8-Flash-Next	Strix Halo		llama.cpp/hip (speculative decode, mtp)	55.3 tok/s	262144	Extracted, source-linked	source
Qwen3.8-Flash-Next	Strix Halo		strixite/hip (speculative decode, mtp)	48.5 tok/s (prefill 1370)	512000	Extracted, source-linked	source
Qwen3.8-Flash-Next	Strix Halo	W4B	halogen/rocm (speculative decode, mtp)	46 tok/s	32768	Extracted, source-linked	source
Qwen3.8-Flash-Next	DGX Spark	NVFP4	vllm (speculative decode, mtp-2)	44.2 tok/s	262144	Extracted, source-linked	source
Qwen3.8-Flash-Next	Strix Halo	UD-Q4_K_XL	llama.cpp/rocm (speculative decode, mtp)	43.9 tok/s (prefill 1207)		Extracted, source-linked	source
Qwen3.8-Flash-Next	Mac Max	UD-Q4_K_XL	llama.cpp	41.4 tok/s (prefill 789)	530	Extracted, source-linked	source
Qwen3.8-Flash-Next	Strix Halo	NVFP4		41.3 tok/s (prefill 273)		Extracted, source-linked	source
Qwen3.8-Flash-Next	DGX Spark	NVFP4	vllm (speculative decode, mtp-3)	39.8 tok/s		Extracted, source-linked	source
Qwen3.8-Flash-Next	Strix Halo	W4B	halogen/rocm (baseline)	37.6 tok/s	1500	Extracted, source-linked	source
Qwen3.8-Flash-Next	DGX Spark	NVFP4	atlas	36.7 tok/s		Extracted, source-linked	source
Qwen3.8-Flash-Next	Strix Halo	IQ3_XXS	llama.cpp (speculative decode, mtp)	35.7 tok/s	156000	Extracted, source-linked	source
Qwen3.8-Flash-Next	DGX Spark	NVFP4	vllm	35.1 tok/s		Structured table	source
Qwen3.8-Flash-Next	Strix Halo	UD-Q4_K_XL	gufo (speculative decode, mtp)	33.4 tok/s (prefill 1689.4)		Extracted, source-linked	source
Qwen3.8-Flash-Next	DGX Spark	NVFP4	vllm (speculative decode, mtp-1)	27.6 tok/s		Extracted, source-linked	source
Qwen3.8-Flash-Next	DGX Spark	NVFP4	vllm (baseline)	27.3 tok/s	262144	Extracted, source-linked	source
Qwen3.8-Flash-Next	Strix Halo	IQ4_XS	llama.cpp/vulkan	27.2 tok/s (prefill 394.7)		Structured table	source
Qwen3.8-Flash-Next	DGX Spark	NVFP4	vllm	26.7 tok/s	1000000	Extracted, source-linked	source
Qwen3.8-Flash-Next	Strix Halo	Q4_K_XL	llama.cpp/vulkan	24 tok/s		Extracted, source-linked	source
Qwen3.8-Flash-Next	Strix Halo	IQ3_XXS	llama.cpp (baseline)	23.5 tok/s	156000	Extracted, source-linked	source
Qwen3.8-Flash-Next	DGX Spark	NVFP4		22.3 tok/s (prefill 1117.9)		Extracted, source-linked	source

Qwen3.8-Flash-Next on GPU cards, ranked on their own

GPU cards run the model from their own VRAM: very fast on what fits, but most cannot load what a 128 GB AI box can. Ranked on their own, never against the boxes.

No single-stream configs recorded yet.

More: all configs · hardware · news · contributors · methodology · llms.txt · llms-full.txt · API (OpenAPI) · JSON snapshot. Built by Altronis, private on-prem AI, Singapore.