Reasoning on or off: Qwen gains 33 points, big Gemma 10, the small Gemmas nothing

Most local models can be run in two modes: with a reasoning phase before the answer, or straight to the answer. The reasoning phase costs time and tokens. Does it buy better writing? We ran 21 local builds both ways on the same English prompts and scored both versions in the same blind duels. Each arrow below starts at the score without reasoning and ends at the score with it.

Twenty-one builds, one switch: the arrow points right for nineteen of them

What the reasoning switch changes

Score with reasoning off (start of the arrow) and on (tip of the arrow), 21 local builds sorted by gain.

• Qwen
• Gemma
• Reasoning off
• Reasoning on

View as a data table

Model	Reasoning off	Reasoning on	Gain
Qwen3.8 27B Q4_K_M	18.8	64.1	+45.3
Qwen3.8 27B Q8_0	20.5	61.3	+40.8
Qwen3.6 27B Q8_0	30.0	62.5	+32.5
Qwen3.8 27B Q6_K	27.9	59.0	+31.1
Qwen3.6 35B A3B Q8_0	24.0	54.4	+30.4
Gemma 4 12B Q6_K	26.1	48.6	+22.5
Qwen3.6 35B A3B Q4_K_M	25.2	45.2	+20.0
Gemma 4 12B QAT-Q4_0	31.4	46.4	+15.0
Gemma 4 26B A4B Q8_0	32.0	46.0	+14.0
Gemma 4 26B A4B Q6_K	46.4	57.4	+11.0
Gemma 4 31B Q6_K	45.3	55.7	+10.4
Gemma 4 31B Q8_0	44.8	54.9	+10.1
Gemma 4 E2B Q8_0	12.5	21.3	+8.8
Gemma 4 E2B Q6_K	6.3	13.9	+7.6
Gemma 4 12B Q8_0	39.9	44.2	+4.3
Gemma 4 E2B Q4_K_M	8.6	12.9	+4.3
Gemma 4 31B QAT-Q4_0	51.3	54.3	+3.0
Gemma 4 E4B Q6_K	16.2	18.6	+2.4
Gemma 4 26B A4B QAT-Q4_0	42.2	44.1	+1.9
Gemma 4 E4B Q8_0	15.3	14.6	−0.7
Gemma 4 E4B Q4_K_M	18.6	11.8	−6.8

Nineteen builds gain, two lose. But the length of the arrows is the story: the Qwen builds gain 20 to 45 points, the larger Gemma 4 builds 2 to 23, the small Gemmas hover around zero.

Qwen gains 33 points on average, big Gemma 10, small Gemma 3

What reasoning adds, family by family

Average score gain when reasoning is switched on, over the builds of each family.

View as a data table

	points	n
Qwen 3.6 / 3.8	+33.4	6
Gemma 4, 12B and larger	+10.2	9
Gemma 4 E2B / E4B	+2.6	6

The families behave so differently that averaging across them would hide everything. Qwen3.6 and Qwen3.8 gain 33.4 points on average; Gemma 4 at 12B and above gains 10.2; Gemma 4 E2B and E4B gain 2.6, with two builds losing.

Without reasoning, Qwen3.8 27B is a weak model; with it, the best local Qwen

Qwen3.8 27B Q4_K_M scores 18.8 without reasoning and 64.1 with it, a gain of 45 points, the largest in the table. Its Q8_0 build gains 41, Qwen3.6 27B Q8_0 gains 32. Whatever these models do in their reasoning phase, their English prose depends on it. Run a Qwen without reasoning and you are ranking a different, much weaker model.

Gemma 4 E4B loses points when it thinks

At the other end, Gemma 4 E4B Q4_K_M scores 6.8 points lower with reasoning than without, and E4B Q8_0 loses 0.7. Gemma 4 31B and 26B gain between 2 and 14 points depending on the build, worth having but not decisive. For the small Gemmas, reasoning is time spent for nothing.

The same experiment in French, German and Spanish

The pattern holds. Qwen gains 26 to 33 points on average in every language; the larger Gemma 4 builds gain 5 to 10; the small Gemmas gain nothing and lose in French and German. The one exception worth knowing: in French and German, several Gemma 4 12B builds lose 5 to 11 points with reasoning on.

Conclusion

Reasoning is not a free upgrade and not a gimmick; it depends on the family. Qwen needs it: without it, Qwen3.8 27B is a bottom-third model, with it a top-local one. Gemma 4 at 12B and above gains a few points, the small Gemmas none. Switch it on for Qwen, leave it off for Gemma E2B and E4B, and accept two to four times slower generation whenever it is on.