Benchmarks

Name: DiggTop benchmarks
Creator: DiggTop

Every run we hold, attributed per row. Filter by what you own or what you want to run — model size is the first thing to look at, because a 350M model and a 100B model are not the same measurement.

sourced — engine + version, quantization, context, a named GPU and a link to the author; we have not re-run it self-measured — DiggTop's own tests partial — missing a required field; the badge says which, and it never counts in averages methodology · submit a run

491runs recorded

359sourced

132partial

74GPUs

172models

Fastest GPURadeon AI Pro R9700969.4 tok/s via Qwen3.6-35B-A3B · 36B Most measuredDGX Spark48 verified runs Newest verifiedDGX Spark x476.3 tok/s · DeepSeek-V4.1-Flash-1017

Model sizeAll sizes491 ≤2B29 3–8B45 9–40B266 41–100B4 >100B144 Unknown3

Runs 1–100 of 491 · 359 sourced · sorted by date added DESC · flip · updated Oct 17, 2026

Top GPUs (sourced best)

Radeon AI Pro R9700via Qwen3.6-35B-A3B · 36B969.4
NVIDIA H200 SXM x4via MiniMax-M2.7 · 229B495.9
DGX Spark x2via Qwen3.8-Flash-Next-hibrid48 · 180B474.0
RTX 5090via NVIDIA-Nemotron-3.5-Lightning-30B-A3B · 32B446.7
Tesla V100 x4via Qwen3.8-27B · 28B391.0

Best sourced single-stream run per GPU on models of 7B parameters and up, with the model that produced it.

Top models (sourced best)

Qwen3.6-35B-A3Bon Radeon AI Pro R9700 · 36B969.4
MiniMax-M2.7on NVIDIA H200 SXM x4 · 229B495.9
Qwen3.8-Flash-Next-hibrid48on DGX Spark x2 · 180B474.0
NVIDIA-Nemotron-3.5-Lightning-30B-A3Bon RTX 5090 · 32B446.7
Qwen3.8-27Bon RTX 5090 · 28B435.1

Best sourced single-stream run per model, with the GPU that produced it. Compare like-for-like pairs on /compare.

#	GPU	Model	Size	Quant	Engine	Ctx	tok/s	prompt tok/s	VRAM	Added ↓	Source
1	DGX Spark x4 · 128 GB	DeepSeek-V4.1-Flash-1017	552B	MXFP4+MXFP8	vllm dsv41-feat (commit e47aa780b)	300K	76.3 sourced	—	128 GB	2d	src
2	DGX Spark x3 · 128 GB	DeepSeek-V4.1-Flash-1017	552B	EXL3 3.5bpw	vllm bot-lab-21's EXL3 build + drowzeys' …	300K	152.9 sourced	—	128 GB	2d	src
3	DGX Spark x4 · 128 GB	DeepSeek-V4.1-Flash-1017	552B	EXL3 3.5bpw	vllm bot-lab-21's EXL3 build + drowzeys' …	300K	190.0 sourced	—	128 GB	2d	src
4	DGX Spark · 128 GB	NVIDIA-Nemotron-3.5-Super-VL-120B-A12B	120B	NVFP4	vllm vllm/vllm-openai@sha256:cb9bc64f… …	256K	48.0 sourced	—	128 GB	3d	src
5	DGX Spark x2 · 128 GB	DeepSeek-V4.1-Flash	552B	EXL3 2.7bpw (per-expert)	vllm upstream vLLM with 21 patches	240K	141.0 sourced	—	128 GB	5d	src
6	Mac Studio (M4 Max) · 64 GB	Qwen3.6-35B-A3B	36B	Q4_K_XL	llama.cpp 781dbc5	400	207.6 sourced	—	64 GB	5d	src
7	Mac Studio (M4 Max) · 64 GB	Qwen3.6-35B-A3B	36B	Q4_K_XL	llama.cpp 781dbc5	4K	220.2 sourced	—	64 GB	5d	src
8	Mac Studio (M4 Max) · 64 GB	Qwen3.6-35B-A3B	36B	Q4_K_XL	llama.cpp 781dbc5	1K	226.9 sourced	—	64 GB	5d	src
9	M6 Mac mini	Ternary Bonsai 2 27B	27B	2-bit ternary LUT	mlx 26.10.1	8K	36.3 sourced	—	—	6d	src
10	M6 Mac mini	Ternary Bonsai 2 27B	27B	2-bit ternary LUT	anemll forge 6c572a8	8K	52.5 sourced	—	—	6d	src
11	RTX 3090 x2 · 24 GB	MiniCPM-V 4.7-35B-A3B	35B	int4	vllm 0.31.1rc1.dev296 (nightly fe4df3d)	256K	172.2 sourced	—	24 GB	7d	src
12	DGX Spark x2 · 128 GB	GLM-5.3-Flash	321B	W4A16 AutoRound	vllm v0.31.0	125K	38.5 sourced	—	128 GB	7d	src
13	DGX Spark x2 · 128 GB	GLM-5.3-Flash	321B	W4A16 AutoRound	vllm v0.31.0	1K	39.4 sourced	—	128 GB	7d	src
14	DGX Spark x2 · 128 GB	GLM-5.3-Flash	321B	W4A16 AutoRound	vllm v0.31.0	400K	43.1 sourced	—	128 GB	7d	src
15	DGX Spark x4 · 128 GB	GLM-5.3-Flash	321B	NVFP4	vllm knapcio stack	500K	115.0 sourced	—	128 GB	7d	src
16	DGX Spark x4 · 128 GB	GLM-5.3-Flash	321B	EXL3 4bpw	vllm knapcio stack	500K	115.0 sourced	—	128 GB	7d	src
17	DGX Spark x4 · 128 GB	GLM-5.3-Flash	321B	EXL3 4bpw	tensorfold v0.6.0	500K	124.0 sourced	—	128 GB	7d	src
18	Radeon AI Pro R9700 · 32 GB	Qwen 3.8 Flash-Next Base	125B	IQ3_S	xestrata xe0.1.42	125K	57.0 sourced	119.4	32 GB	8d	src
19	Mac Studio (Mac16,9) with Apple M4 Max · 64 GB	Qwen3.8-Flash-Next	180B	IQ3_XXS + MTP Q8_0	llama.cpp 781dbc5ac98921dbdb5e5b2ec5b7a50960e9…	1K	110.3 sourced	—	64 GB	8d	src
20	Tesla V100 PCIe 16GB · 16 GB	Qwen3.8-Flash-Next	180B	Q2_0	strata v0.1.42	8K	40.7 sourced	—	16 GB	8d	src
21	DGX Spark · 128 GB	Qwen3.8-Flash-Next	180B	4.05bpw_h6_ng6	tensorfold 0.6.5	40K	75.7 sourced	—	128 GB	9d	src
22	DGX Spark · 128 GB	Underdog-Saluki-27B	27B	IQ2-mix	llama.cpp b11459 (f498f864f)	256K	22.6 sourced	—	128 GB	11d	src
23	DGX Spark · 128 GB	CYBER-FROST-3.8	180B	EXL3	exllamav2 047ce72	400	60.7 sourced	—	128 GB	11d	src
24	Intel Arc Pro B70 · 32 GB	Qwen3.8-Flash-Next	180B	EXL3 3.05 bpw	sglang 24c872759256	32K	23.9 sourced	1106.0	32 GB	13d	src
25	Intel Arc Pro B70 · 32 GB	Qwen3.8-Flash-Next	180B	EXL3 3.05 bpw	sglang 24c872759256	8K	24.0 sourced	1056.0	32 GB	13d	src
26	DGX Spark x2 · 128 GB	Qwen3.8-Flash-Next	180B	INT4-AutoRound	vllm 0.30.0 with patches (v5.2 image, sha…	125K	93.0 sourced	—	128 GB	13d	src
27	RTX 3090 · 24 GB	DeepSeek-V4.1-Flash	552B	EXL3 3.0 bpw	vllm 0.13 (lazymio/vllm-backport, DeepSee…	256K	26.0 sourced	—	24 GB	14d	src
28	DGX Spark x2 · 128 GB	DeepSeek-V4.1-Flash	552B	EXL3 2.9bpw	tensorfold bd0024d	125K	89.2 sourced	—	128 GB	14d	src
29	DGX Spark x2 · 128 GB	DeepSeek-V4.1-Flash	552B	EXL3 2.9bpw	tensorfold bd0024d	384	101.1 sourced	—	128 GB	14d	src
30	GT 430	OpenLLaMA-3B	3B	Q4_0	llama.cpp 2e6cd4b	512	2.4 sourced	—	—	15d	src
31	RTX 5070 · 12 GB	Qwen3.8-Flash-Next	180B	Q2_0	strata 0.1.26	32K	81.8 sourced	—	12 GB	15d	src
32	Mac Studio M3 Ultra	GLM-5.3-Flash	321B	4-bit	tensorfold 0.6.2	256	78.0 sourced	—	—	16d	src
33	RTX 3090 · 24 GB	GLM-5.3-Flash	321B	EXL3 3.05 bpw	exllamav2 v1.5.1	128K	28.1 sourced	—	24 GB	18d	src
34	DGX Spark x2 · 128 GB	GLM-5.3-Flash	321B	EXL3 4-bit	vllm Entrpi/vllm-glm-5.3-flash-spark	3700	27.4 sourced	1369.0	128 GB	19d	src
35	M4 Max 128GB	Qwen3.8-27B	28B	4-bit	mlx 0.3.4 → 0.3.6.3	48800	134.0 sourced	—	—	20d	src
36	DGX Spark x2 · 128 GB	GLM-5.3-Flash	321B	NVFP4	vllm glm53-sm121-v13	320K	30.0 sourced	—	128 GB	21d	src
37	DGX Spark x2 · 128 GB	Qwen3.8-Flash-Next-hibrid48	180B	NVFP4	vllm 0.30	250K	81.6 sourced	3094.0	128 GB	21d	src
38	DGX Spark x2 · 128 GB	Qwen3.8-Flash-Next-hibrid48	180B	NVFP4	vllm 0.30	125K	89.8 sourced	3233.0	128 GB	21d	src
39	DGX Spark x2 · 128 GB	Qwen3.8-Flash-Next-hibrid48	180B	NVFP4	vllm 0.30	1K	99.0 sourced	2543.0	128 GB	21d	src
40	CMP 170HX x4 · 64 GB	GLM-5.3-Flash	321B	W4A16	vllm 0ed7d3e7f3	256K	264.6 sourced	—	64 GB	24d	src
41	Tesla V100 x2 · 32 GB	Qwen3.8-Flash-Next	180B	IQ3XXS	llama.cpp b11030-mix-5ff778e	4K	96.1 sourced	—	32 GB	26d	src
42	Tesla V100 x2 · 32 GB	Qwen3.8-Flash-Next	180B	IQ3XXS	llama.cpp b11030-mix-5ff778e	512	87.5 sourced	—	32 GB	26d	src
43	Tesla V100 x2 · 32 GB	Qwen3.8-Flash-Next	180B	IQ3XXS	llama.cpp b11030-mix-5ff778e	131K	80.9 sourced	—	32 GB	26d	src
44	DGX Spark x4 · 128 GB	DeepSeek-V4.1-Flash	552B	EXL3	vllm vllm-exl3	1000K	30.2 sourced	—	128 GB	26d	src
45	CMP 170HX x2 · 64 GB	Qwen3.8-Flash-Next	180B	W4A16-FP8PLE	vllm 0.29.0	262052	186.9 sourced	—	64 GB	27d	src
46	RTX 3090 · 24 GB	Qwen3.8-27B	28B	EXL3 4.00bpw	buun-llama-cpp 1d0f493c73817176f6953069b7c10211de95…	150K	27.8 sourced	525.3	24 GB	28d	src
47	RTX 3090 · 24 GB	Qwen3.8-27B	28B	EXL3 4.00bpw	buun-llama-cpp 1d0f493c73817176f6953069b7c10211de95…	256K	149.2 sourced	—	24 GB	28d	src
48	DGX Spark · 128 GB	Qwen 3.8 Flash	125B	EXL3 3.05bpw	exllamav2 523ecd3	2K	86.8 sourced	—	128 GB	29d	src
49	DGX Spark · 128 GB	Qwen 3.8 Flash	125B	EXL3 3.05bpw	exllamav2 523ecd3	35K	58.0 sourced	980.0	128 GB	29d	src
50	DGX Spark · 128 GB	Qwen 3.8 Flash	125B	EXL3 3.05bpw	exllamav2 523ecd3	240K	55.8 sourced	—	128 GB	29d	src
51	DGX Spark · 128 GB	Qwen 3.8 Flash	125B	EXL3 3.05bpw	exllamav2 523ecd3	110K	66.0 sourced	—	128 GB	29d	src
52	DGX Spark · 128 GB	Qwen 3.8 Flash	125B	EXL3 3.05bpw	exllamav2 523ecd3	64K	67.4 sourced	—	128 GB	29d	src
53	DGX Spark · 128 GB	Qwen 3.8 Flash	125B	EXL3 3.05bpw	exllamav2 523ecd3	32K	55.6 sourced	—	128 GB	29d	src
54	DGX Spark · 128 GB	Qwen 3.8 Flash	125B	EXL3 3.05bpw	exllamav2 523ecd3	8K	51.5 sourced	—	128 GB	29d	src
55	DGX Spark · 128 GB	Qwen 3.8 Flash	125B	EXL3 3.05bpw	exllamav2 523ecd3	600	71.7 sourced	—	128 GB	29d	src
56	DGX Spark · 128 GB	Qwen3.8-Flash-Next	180B	NVFP4	vllm 8a728663	256K	43.9 sourced	—	128 GB	29d	src
57	DGX Spark · 128 GB	Qwen3.8-Flash-Next	180B	EXL3 3.05bpw	exllamav2 523ecd3	75K	67.6 sourced	—	128 GB	Sep 18	src
58	DGX Spark · 128 GB	Qwen3.8-Flash-Next	180B	EXL3 3.05bpw	exllamav2 523ecd3	32K	59.0 sourced	—	128 GB	Sep 18	src
59	DGX Spark · 128 GB	Qwen3.8-Flash-Next	180B	EXL3 3.05bpw	exllamav2 523ecd3	1K	56.0 sourced	—	128 GB	Sep 18	src
60	DGX Spark · 128 GB	Qwen3.8-Flash-Next	180B	EXL3	exllamav2 523ecd3	125K	69.5 sourced	—	128 GB	Sep 18	src
61	DGX Spark · 128 GB	Qwen3.8-Flash-Next	180B	EXL3	exllamav2 523ecd3	4K	69.8 sourced	—	128 GB	Sep 18	src
62	DGX Spark · 128 GB	Qwen3.8-Flash-Next	180B	EXL3	exllamav2 523ecd3	240K	72.0 sourced	—	128 GB	Sep 18	src
63	RTX 3090 · 24 GB	Qwen3.8-Flash-Next	180B	EXL3 2.50bpw	exllamav2 v1.5.0-compatible build with MoE CPU…	175K	20.9 sourced	—	24 GB	Sep 17	src
64	RTX 3090 · 24 GB	Qwen3.8-Flash-Next	180B	EXL3 2.50bpw	exllamav2 v1.5.0-compatible build with MoE CPU…	256K	38.6 sourced	—	24 GB	Sep 17	src
65	RTX 3090 · 24 GB	Qwen3.8-27B	28B	EXL3 4.00bpw	exllamav2 v1.5.0	150K	24.5 sourced	—	24 GB	Sep 16	src
66	RTX 3090 · 24 GB	Qwen3.8-27B	28B	EXL3 4.00bpw	exllamav2 v1.5.0	8K	152.4 sourced	—	24 GB	Sep 16	src
67	DGX Spark · 128 GB	Qwen3.8-Flash-Next	180B	NVFP4	sglang 8874c51a	256K	55.0 sourced	—	128 GB	Sep 16	src
68	RTX 4090 · 24 GB	Qwen3.8-27B	28B	int4 (output head only)	vllm pinned version (exact version not sp…	38K	141.2 sourced	—	24 GB	Sep 15	src
69	DGX Spark · 128 GB	GLM-5.3-Flash	321B	EXL3 2.05 bpw	vllm sha256:905c02933be6021301db2dc284e24…	256K	64.1 sourced	—	128 GB	Sep 14	src
70	DGX Spark · 128 GB	GLM-5.3-Flash	321B	EXL3 2.05 bpw	vllm sha256:905c02933be6021301db2dc284e24…	40K	35.5 sourced	—	128 GB	Sep 14	src
71	DGX Spark x2 · 128 GB	DeepSeek-V4.1-Flash	552B	EXL3 2.9bpw	exl3 v1.4.2	559K	30.5 sourced	—	128 GB	Sep 14	src
72	DGX Spark x2 · 128 GB	DeepSeek-V4.1-Flash	552B	EXL3 2.9bpw	exl3 v1.4.2	268K	32.6 sourced	—	128 GB	Sep 14	src
73	DGX Spark x2 · 128 GB	DeepSeek-V4.1-Flash	552B	EXL3 2.9bpw	exl3 v1.4.2	67K	35.0 sourced	—	128 GB	Sep 14	src
74	DGX Spark x2 · 128 GB	DeepSeek-V4.1-Flash	552B	EXL3 2.9bpw	exl3 v1.4.2	17K	32.4 sourced	—	128 GB	Sep 14	src
75	DGX Spark x2 · 128 GB	DeepSeek-V4.1-Flash	552B	EXL3 2.9bpw	exl3 v1.4.2	4K	154.0 sourced	—	128 GB	Sep 14	src
76	DGX Spark x2 · 128 GB	DeepSeek-V4.1-Flash	552B	EXL3 2.9bpw	exl3 v1.4.2	2K	74.6 sourced	—	128 GB	Sep 14	src
77	DGX Spark x2 · 128 GB	DeepSeek-V4.1-Flash	552B	EXL3 2.9bpw	exl3 v1.4.2	1K	45.5 sourced	—	128 GB	Sep 14	src
78	DGX Spark · 128 GB	Qwen3.8-Flash-Next-hibrid48	180B	NVFP4 (output head)	vllm 0.29.0	262K	60.0 sourced	—	128 GB	Sep 14	src
79	DGX Spark x2 · 128 GB	Qwen3.8-Flash-Next-hibrid48	180B	NVFP4 (output head)	vllm 0.29	256K	474.0 sourced	—	128 GB	Sep 13	src
80	DGX Spark · 128 GB	DeepSeek-V4.1-Flash	552B	FP4	custom v41_engine.py v0.1.0-wip (2026-09-10)	32K	47.0 sourced	—	128 GB	Sep 13	src
81	DGX Spark · 128 GB	DeepSeek-V4.1-Flash	552B	FP4	custom v41_engine.py v0.2.0-wip (commits bd24743..22bd9a8…	8K	16.9 sourced	—	128 GB	Sep 13	src
82	DGX Spark · 128 GB	DeepSeek-V4.1-Flash	552B	FP4	custom v41_engine.py v0.3.0-wip (commit 94a96a6..)	32K	20.9 sourced	—	128 GB	Sep 13	src
83	DGX Spark x4 · 128 GB	DeepSeek-V4.1-Flash	552B	EXL3 3.5bpw	vllm dsv41-feat commit e47aa780b	300K	73.8 sourced	—	128 GB	Sep 11	src
84	DGX Spark x4 · 128 GB	DeepSeek-V4.1-Flash	552B	EXL3 3.5bpw	vllm e47aa780b	300K	82.0 sourced	—	128 GB	Sep 11	src
85	DGX Spark x4 · 128 GB	DeepSeek-V4.1-Flash	552B	MXFP4	vllm e47aa780b	300K	38.0 sourced	1194.0	128 GB	Sep 11	src
86	DGX Spark x3 · 128 GB	DeepSeek-V4.1-Flash	552B	EXL3 3.5bpw	vllm e47aa780b	300K	46.0 sourced	—	128 GB	Sep 11	src
87	RTX 3090 x2 · 24 GB	Qwen3.8-27B	28B	AutoRound INT4 + fp8 KV	sglang 0.5.19	262K	219.8 sourced	—	24 GB	Sep 11	src
88	DGX Spark · 128 GB	DeepSeek-V4.1-Flash	552B	FP4	custom triton-based engine (v41_ 22bd9a8 (v0.2.0-wip baseline)	8K	13.0 sourced	—	128 GB	Sep 11	src
89	DGX Spark · 128 GB	DeepSeek-V4.1-Flash	552B	FP4	custom triton-based engine (v41_ 9f172fb (v0.2.0-wip addendum 2.6)	8K	15.4 sourced	—	128 GB	Sep 11	src
90	DGX Spark · 128 GB	DeepSeek-V4.1-Flash	552B	FP4	custom triton-based engine (v41_ bd24743..22bd9a8+ (v0.2.0-wip), then…	32K	20.9 sourced	—	128 GB	Sep 11	src
91	DGX Spark · 128 GB	DeepSeek-V4.1-Flash	552B	FP4	custom v41_engine.py v0.3.0-wip (3.3 addendum)	32K	22.1 sourced	—	128 GB	Sep 11	src
92	DGX Spark · 128 GB	DeepSeek-V4.1-Flash	552B	FP4	custom v41_engine.py v0.3.0-wip (3.2 addendum)	32K	20.9 sourced	—	128 GB	Sep 11	src
93	DGX Spark · 128 GB	DeepSeek-V4.1-Flash	552B	FP4	custom v41_engine.py v0.2.0-wip (2.9 addendum)	8K	16.9 sourced	—	128 GB	Sep 11	src
94	DGX Spark · 128 GB	DeepSeek-V4.1-Flash	552B	FP4	custom v41_engine.py v0.2.0-wip (9f172fb)	8K	15.4 sourced	—	128 GB	Sep 11	src
95	DGX Spark · 128 GB	DeepSeek-V4.1-Flash	552B	FP4	custom v41_engine.py v0.2.0-wip (bd24743..22bd9a8+)	8K	12.9 sourced	—	128 GB	Sep 11	src
96	DGX Spark · 128 GB	DeepSeek-V4.1-Flash	552B	FP4	custom v41_engine.py v0.1.0-wip	32K	2.7 sourced	—	128 GB	Sep 11	src
97	DGX Spark x2 · 128 GB	GLM-5.3-Flash	321B	NVFP4	vllm 0.28.1rc1.dev580	126720	46.3 sourced	—	128 GB	Sep 10	src
98	RTX PRO 6000 Blackwell x4 · 384 GB	DeepSeek-V4.1-Flash	552B	FP4+FP8 native	sglang lmsysorg/sglang c4ca6511 + nvme/DDR5…	500K	212.4 sourced	5844.5	384 GB	Sep 10	src
99	RTX PRO 6000 Blackwell x4 · 384 GB	DeepSeek-V4.1-Flash	552B	FP4+FP8 native	sglang lmsysorg/sglang c4ca6511 + nvme/DDR5…	400K	220.2 sourced	6081.4	384 GB	Sep 10	src
100	RTX PRO 6000 Blackwell x4 · 384 GB	DeepSeek-V4.1-Flash	552B	FP4+FP8 native	sglang lmsysorg/sglang c4ca6511 + nvme/DDR5…	200K	202.2 sourced	6802.6	384 GB	Sep 10	src