Benchmarks

Depth	Gufo pp	llama.cpp pp	Gain	Gufo tg	llama.cpp tg	Gain
0	1700.52	489.59	+247.3%	25.95	22.20	+16.9%
4,096	1618.60	455.24	+255.5%	25.84	21.21	+21.8%
8,192	1611.53	428.85	+275.8%	25.96	20.40	+27.3%
12,288	1602.99	400.21	+300.5%	25.94	19.64	+32.1%
16,384	1597.41	375.89	+325.0%	25.90	18.95	+36.7%
32,768	1585.98	301.31	+426.4%	25.78	16.54	+55.9%
65,536	1509.10	221.94	+580.0%	25.40	11.62	+118.6%
131,072	1478.66	144.78	+921.3%	24.92	7.98	+212.3%

Depth

Gufo pp

llama.cpp pp

Gain

Gufo tg

llama.cpp tg

Gain

1700.52

489.59

+247.3%

25.95

22.20

+16.9%

4,096

1618.60

455.24

+255.5%

25.84

21.21

+21.8%

8,192

1611.53

428.85

+275.8%

25.96

20.40

+27.3%

12,288

1602.99

400.21

+300.5%

25.94

19.64

+32.1%

16,384

1597.41

375.89

+325.0%

25.90

18.95

+36.7%

32,768

1585.98

301.31

+426.4%

25.78

16.54

+55.9%

65,536

1509.10

221.94

+580.0%

25.40

11.62

+118.6%

131,072

1478.66

144.78

+921.3%

24.92

7.98

+212.3%

Depth	Gufo pp	llama.cpp pp	Gufo tg mixed	llama.cpp tg mixed	Gufo tg repetitive	llama.cpp tg repetitive
0	1642.85	468.76	31.32	31.73	60.39	48.12
4,096	1575.60	423.79	32.48	34.48	57.66	45.71
8,192	1573.54	394.98	32.38	32.20	49.28	44.31
12,288	1559.07	365.91	34.68	32.32	42.15	43.53
16,384	1547.52	341.98	35.09	32.07	50.58	43.66
32,768	1541.35	277.61	34.20	25.41	44.54	39.28
65,536	1466.98	208.04	33.49	19.65	44.45	27.75
131,072	1436.89	135.42	33.07	14.63	44.26	20.96

Users	Gufo AR	llama.cpp AR	Gufo MTP mixed	llama.cpp MTP mixed	Gufo MTP repetitive	llama.cpp MTP repetitive
1	25.84	22.34	31.28	31.77	60.10	46.93
2	45.77	37.08	52.22	42.79	84.52	52.57
4	76.33	54.82	76.16	48.78	131.10	48.23
6	95.75	65.34	96.01	52.59	145.38	48.51
8	108.68	68.79	107.77	61.92	162.98	58.13

Measurement	Gufo	llama.cpp	Gain
Ready (s)	15.45	117.83	+662.7%
Peak memory, pp2048 + tg128 (GiB)	85.56	84.84	-0.8%
Peak memory, 16K prefix, pp4096 + tg128 (GiB)	86.71	85.31	-1.6%

Prefill across context depth

Decode across context depth

Single user, autoregressive, Q4

Approximately pp2048 / tg128.

Depth	Gufo pp	llama.cpp pp	Gain	Gufo tg	llama.cpp tg	Gain
0	656.33	357.54	+83.6%	12.37	12.06	+2.6%
4,096	633.16	330.42	+91.6%	12.21	11.88	+2.8%
8,192	615.06	316.09	+94.6%	12.03	11.73	+2.6%
12,288	594.61	302.54	+96.5%	11.88	11.59	+2.5%
16,384	575.50	290.23	+98.3%	11.72	11.44	+2.4%
32,768	505.07	246.97	+104.5%	11.13	10.89	+2.2%
65,536	386.92	193.61	+99.8%	10.13	9.93	+2.0%
131,072	287.12	135.95	+111.2%	8.52	8.46	+0.7%

Prefill across context depth

Decode across context depth

Single user, autoregressive, Q8

Decode is bandwidth bound at this precision: both engines land on the same rate.

Depth	Gufo pp	llama.cpp pp	Gain	Gufo tg	llama.cpp tg	Gain
0	485.26	337.08	+44.0%	7.20	7.20	+0.0%
4,096	470.80	317.96	+48.1%	7.14	7.14	+0.0%
8,192	450.84	306.22	+47.2%	7.08	7.08	+0.0%
12,288	436.37	289.19	+50.9%	7.03	7.02	+0.1%
16,384	423.19	276.91	+52.8%	6.97	6.97	+0.0%
32,768	381.43	236.32	+61.4%	6.76	6.76	+0.0%
65,536	311.47	186.67	+66.9%	6.38	6.38	+0.0%
131,072	248.60	133.34	+86.4%	5.71	5.74	-0.5%

Decode at depth 0 by execution mode

DFlash2 decode across context depth, repetitive output

Single user, DFlash2, Q4

pp is the highest measured rate per engine and depth across mixed and repetitive text.

Depth	Gufo pp	llama.cpp pp	Gufo tg mixed	llama.cpp tg mixed	Gufo tg repetitive	llama.cpp tg repetitive
0	607.27	347.51	26.13	23.40	63.12	35.86
4,096	559.04	316.58	27.06	23.03	52.24	34.47
8,192	537.00	302.47	26.77	21.73	61.55	34.73
12,288	520.81	289.98	23.92	23.05	53.18	34.19
16,384	504.59	278.53	23.78	21.84	49.67	33.75
32,768	448.96	239.21	19.25	18.88	47.76	31.75
65,536	347.29	190.92	17.70	18.53	29.44	26.33
131,072	274.07	133.37	14.70	14.58	26.83	22.93

Decode at depth 0 by execution mode

DFlash2 decode across context depth, repetitive output

Single user, DFlash2, Q8

Mixed-text decode falls behind the reference at the two deepest points; repetitive output stays well ahead.

Depth	Gufo pp	llama.cpp pp	Gufo tg mixed	llama.cpp tg mixed	Gufo tg repetitive	llama.cpp tg repetitive
0	457.05	333.85	16.08	15.00	46.84	23.83
4,096	423.09	310.23	16.36	16.12	40.51	23.59
8,192	408.06	292.86	15.57	15.27	41.53	23.21
12,288	401.07	280.40	14.99	13.21	40.45	22.10
16,384	390.69	269.03	14.28	13.83	39.39	22.56
32,768	355.59	231.89	14.41	14.44	35.87	21.16
65,536	293.81	185.73	10.87	12.62	25.45	19.03
131,072	237.24	131.80	8.87	11.62	23.41	17.23

Summed decode rates by concurrency

Multiple users, Q4

Context 4096 per user, tg128. Throughput sums individual request decode rates, excluding prefill and scheduling.

Users	Gufo AR	llama.cpp AR	Gufo DFlash2 mixed	llama.cpp DFlash2 mixed	Gufo DFlash2 repetitive	llama.cpp DFlash2 repetitive
1	12.42	12.20	36.30	27.15	70.56	37.56
2	23.84	22.58	56.34	40.04	105.51	51.65
4	44.23	38.94	70.87	74.43	110.49	93.38
6	60.99	47.45	79.83	77.19	113.52	97.52
8	74.46	50.14	85.39	71.71	123.00	113.56

Summed decode rates by concurrency

Multiple users, Q8

Mixed prompts and repetitive output, context 4096 per user, up to 128 output tokens. No prompt-cache hits.

Users	Gufo AR	llama.cpp AR	Gufo DFlash2 mixed	llama.cpp DFlash2 mixed	Gufo DFlash2 repetitive	llama.cpp DFlash2 repetitive
1	7.22	7.24	24.65	17.54	51.23	24.09
2	14.45	13.71	42.87	31.64	86.58	41.60
4	27.81	25.32	48.44	43.88	102.98	58.92
6	39.96	34.64	54.77	53.14	109.42	72.40
8	51.62	41.57	65.52	60.64	118.07	86.25

Cold files to HTTP readiness

Peak memory reported by HIP, pp2048 + tg128

Loading and memory

Loading: C1, context capacity 262144, DFlash2. Memory: C1, context capacity 262144, AR.

Measurement	Gufo	llama.cpp	Gain
Ready, Q4 (s)	5.87	23.33	+297.4%
Ready, Q8 (s)	8.17	36.47	+346.4%
Peak memory, Q4, pp2048 + tg128 (GiB)	37.63	36.19	-3.8%
Peak memory, Q4, 16K prefix, pp4096 + tg128 (GiB)	39.88	36.62	-8.2%
Peak memory, Q8, pp2048 + tg128 (GiB)	50.58	48.60	-3.9%
Peak memory, Q8, 16K prefix, pp4096 + tg128 (GiB)	52.83	49.02	-7.2%

Notes

· With greedy decoding, DFlash2 output matches plain decoding. llama.cpp differs in a few settings; the quality report has the details.
· Multi-user numbers add up each request’s decode speed and leave out prefill and queueing.
· The draft doesn’t need to match the main model’s precision: one Q4_K_M draft serves both.

Depth	Gufo pp	ds4 pp	Gain	Gufo tg	ds4 tg	Gain
0	484.62	248.41	+95.1%	16.28	14.83	+9.8%
4,096	399.75	238.93	+67.3%	15.97	14.12	+13.1%
8,192	368.17	223.24	+64.9%	15.82	13.90	+13.8%
12,288	393.95	212.41	+85.5%	15.64	13.78	+13.5%
16,384	358.41	214.97	+66.7%	15.50	13.64	+13.6%
32,768	370.10	205.70	+79.9%	14.87	12.96	+14.7%
65,536	321.28	177.43	+81.1%	13.95	12.00	+16.2%
131,072	279.95	140.04	+99.9%	12.04	10.15	+18.6%

Depth

Gufo pp

ds4 pp

Gain

Gufo tg

ds4 tg

Gain

484.62

248.41

+95.1%

16.28

14.83

+9.8%

4,096

399.75

238.93

+67.3%

15.97

14.12

+13.1%

8,192

368.17

223.24

+64.9%

15.82

13.90

+13.8%

12,288

393.95

212.41

+85.5%

15.64

13.78

+13.5%

16,384

358.41

214.97

+66.7%

15.50

13.64

+13.6%

32,768

370.10

205.70

+79.9%

14.87

12.96

+14.7%

65,536

321.28

177.43

+81.1%

13.95

12.00

+16.2%

131,072

279.95

140.04

+99.9%

12.04

10.15

+18.6%

Depth	Gufo pp	ds4 pp	Gufo tg mixed	Gufo tg repetitive	ds4 tg
0	483.11	N/A	18.25	26.62	N/A
4,096	398.06	N/A	16.01	25.54	N/A
8,192	367.84	N/A	16.48	25.03	N/A
12,288	393.32	N/A	15.79	23.87	N/A
16,384	358.90	N/A	15.52	23.81	N/A
32,768	367.12	N/A	14.98	22.90	N/A
65,536	336.67	N/A	14.05	19.76	N/A
131,072	283.45	N/A	11.95	15.17	N/A

Users	Gufo AR	ds4 AR	Gain	Gufo DSpark mixed	Gufo DSpark repetitive	ds4 DSpark
1	16.27	14.81	+9.9%	18.24	26.59	N/A
2	26.88	14.81	+81.5%	30.80	41.97	N/A
4	41.09	14.89	+176.0%	41.20	47.04	N/A
6	49.21	15.01	+227.8%	46.35	42.04	N/A
8	55.47	14.99	+270.0%	54.74	52.77	N/A

Measurement	Gufo	antirez/ds4	Gain
Ready (s)	16.41	150.87	+819.4%
Peak memory, pp2048 + tg128 (GiB)	93.09	99.96	+7.4%
Peak memory, 16K prefix, pp4096 + tg128 (GiB)	93.88	100.01	+6.5%

Qwen3.8 Flash-Next

Prefill across context depth

Decode across context depth

Single user, autoregressive

Decode at depth 0 by execution mode

MTP decode across context depth, repetitive output

Single user, MTP

Summed decode rates by concurrency

Multiple users

Cold files to HTTP readiness

Loading and memory

Qwen3.8 27B

Prefill across context depth

Decode across context depth

Single user, autoregressive, Q4

Prefill across context depth

Decode across context depth

Single user, autoregressive, Q8

Decode at depth 0 by execution mode

DFlash2 decode across context depth, repetitive output

Single user, DFlash2, Q4

Decode at depth 0 by execution mode

DFlash2 decode across context depth, repetitive output

Single user, DFlash2, Q8

Summed decode rates by concurrency

Multiple users, Q4

Summed decode rates by concurrency

Multiple users, Q8

Cold files to HTTP readiness

Peak memory reported by HIP, pp2048 + tg128

Loading and memory

DeepSeek V4 Flash

Prefill across context depth

Decode across context depth

Single user, autoregressive

Gufo decode across context depth, AR versus DSpark

Single user, DSpark

Summed decode rates by concurrency

Multiple users

Cold files to HTTP readiness

Peak memory reported by HIP

Loading and memory