Parallel · 10 people at once

MTP 3 vs MTP 5: speculative decoding

Validra Assistant and AI Editor, run at both settings on the same machine.

Measured on the live system · Generated 27 October 2026

Keep MTP 3.· Assistant +2.1% on MTP 5.· Editor -3.8%.· First-word wait +30% overall.

Validra Assistantten people at once, four set questions

MTP 3 — speed

94.5

tokens a second per person

MTP 5 — speed

96.4

+2.1% against MTP 3

MTP 3 — first word

0.39s

wait before text appears

MTP 5 — first word

0.68s

+76% against MTP 3

MTP 3 — accepted

62.9%

share of guesses kept

MTP 5 — accepted

49.6%

-13.3 points

AI Editorten people at once, an edit then a generate each

MTP 3 — speed

96.5

tokens a second per person

MTP 5 — speed

92.9

-3.8% against MTP 3

MTP 3 — first word

0.80s

wait before text appears

MTP 5 — first word

0.73s

-9% against MTP 3

MTP 3 — accepted

66.2%

share of guesses kept

MTP 5 — accepted

47.2%

-18.9 points

Speed by run

Figures read from the server's own counters immediately before and after each run.

Tool	Setting	Requests	Words out	Rounds	Words per round	Guesses accepted	Speed per person	First-word wait (s)
Validra Assistant	MTP 3	40	143,300	46,460	16,109	2.89	62.9%	94.5	0.39
Validra Assistant	MTP 5	36	129,901	42,232	12,147	3.48	49.6%	96.4	0.68
AI Editor	MTP 3	23	74,304	21,110	7,077	2.98	66.2%	96.5	0.80
AI Editor	MTP 5	24	90,419	23,984	7,152	3.36	47.2%	92.9	0.73
Both tools together	MTP 3	63	—	67,570	23,186	2.915	63.9%	95.21	0.54
Both tools together	MTP 5	60	—	66,216	19,299	3.434	48.7%	94.97	0.70

Validra Assistant — every run

Ten people at once, all sending at the same moment. Four set questions.

MTP 3

Run	Question	Thinking	Tokens	Model (s)	Total wait (s)	Tok/s in model	Tok/s end to end	Status
1	s0	on	0	0.0	3.0	—	—	passed
2	s1	on	512	7.6	8.3	67.5	61.7	passed
3	s2	on	6,738	86.5	100.4	77.9	67.1	passed
4	s3	on	22,292	311.8	412.2	71.5	54.1	passed
5	s0	off	0	0.0	2.9	—	—	passed
6	s1	off	0	0.0	0.7	—	—	passed
7	s2	off	368	5.5	13.8	67.0	26.7	passed
8	s3	off	3,521	42.6	454.8	82.7	7.7	passed
9	s0	on	0	0.0	2.9	—	—	passed
10	s1	on	512	7.4	462.3	69.3	1.1	passed

MTP 5

Run	Question	Thinking	Tokens	Model (s)	Total wait (s)	Tok/s in model	Tok/s end to end	Status
1	s0	on	0	0.0	4.3	—	—	passed
2	s1	on	0	0.0	0.7	—	—	passed
3	s2	on	4,302	54.4	55.1	79.1	78.1	passed
4	s3	on	20,743	254.0	309.1	81.7	67.1	passed
5	s0	off	0	0.0	4.4	—	—	passed
6	s1	off	0	0.0	0.7	—	—	passed
7	s2	off	673	8.1	317.2	83.2	2.1	passed
8	s3	off	3,981	44.6	361.8	89.3	11.0	passed
9	s0	on	0	0.0	4.3	—	—	passed
10	s1	on	0	0.0	0.6	—	—	passed

“Tok/s in model” is the speed of the last model call; “end to end” is what the person experienced, including queueing. A dash means the run never reached the model.

AI Editor — every run

Ten people at once, two requests each: an edit, then a generate.

MTP 3

Run	Thinking	Edit tokens	Edit (s)	Edit tok/s	Gen tokens	Gen (s)	Gen tok/s	Status
1	on	0	19.3	—	4,826	73.2	83.1	passed
2	off	80	18.0	42.6	0	21.7	—	passed
3	on	492	18.7	155.8	5,245	83.5	76.7	passed
4	off	0	17.7	—	1,995	21.8	299.8	passed
5	on	578	22.0	97.3	3,881	63.0	81.1	passed
6	off	0	17.6	—	0	21.7	—	passed
7	on	0	20.9	—	0	15.2	—	passed
8	off	132	17.1	70.7	0	21.7	—	passed
9	on	0	19.2	—	3,881	63.1	80.9	passed
10	off	0	18.1	—	0	21.7	—	passed

MTP 5

Run	Thinking	Edit tokens	Edit (s)	Edit tok/s	Gen tokens	Gen (s)	Gen tok/s	Status
1	on	0	17.7	—	8,595	70.8	154.2	passed
2	off	40	16.5	27.4	0	20.4	—	passed
3	on	164	17.4	72.7	233	70.8	4.2	passed
4	off	40	16.4	29.1	0	20.6	—	passed
5	on	320	17.9	122.9	2,039	41.8	76.3	passed
6	off	40	16.5	29.5	258	23.7	29.8	passed
7	on	164	18.1	56.8	4,355	72.0	76.5	passed
8	off	40	16.8	28.7	762	20.5	141.2	passed
9	on	162	17.4	68.5	6,514	98.2	78.4	passed
10	off	0	1.2	—	258	20.9	44.1	passed

A dash means the response carried no token count, so no rate could be worked out.