Small Models scoreboard

Each row is one model on one host at one thinking setting. Each question is sent on its own with a fresh context and graded by running or rendering the answer. A cell shows how many takes passed. Thinking “default” means no setting was sent, so the host decides; the reasoning column shows what it did.

Show 18 earlier runs with mixed hosts

Overall

Model	Host	Quant.	Thinking	Mean	AWS Quick Ten	Python Quick Ten	Rust Quick Ten	Tailwind Quick Ten
muse-spark-open	baseten/fp4	fp4	default	99.2%	100.0%	100.0%	96.7%	100.0%
ernie-5.5-flash-preview:free	baidu	unknown	default	97.5%	96.7%	100.0%	93.3%	100.0%
nex-n2.6	nex-agi/fp8	fp8	default	94.2%	93.3%	96.7%	86.7%	100.0%
glm-5.3-flash	together	unknown	default	94.2%	100.0%	90.0%	90.0%	96.7%
nemotron-3-ultra-550b-a55b:free	nvidia	unknown	default	93.3%	90.0%	100.0%	86.7%	96.7%
beam-1.0-501b-a23b	nebius/fp4	fp4	default	91.7%	86.7%	96.7%	83.3%	100.0%
inkling	together	unknown	default	91.7%	76.7%	100.0%	93.3%	96.7%
nemotron-3.5-super-vl	crusoe/bf16	bf16	default	90.0%	80.0%	100.0%	86.7%	93.3%
gemma-4-26b-a4b-it	nextbit/bf16	bf16	default	85.0%	73.3%	93.3%	73.3%	100.0%
gemma-4-26b-a4b-it	akashml/fp4	fp4	default	84.2%	73.3%	96.7%	70.0%	96.7%
ripple-65b-preview	gmicloud/bf16	bf16	default	82.5%	73.3%	93.3%	76.7%	86.7%
gpt-oss-20b	deepinfra/bf16	bf16	default	74.2%	40.0%	96.7%	66.7%	93.3%
gpt-oss-20b	akashml/fp4	fp4	default	72.5%	56.7%	90.0%	53.3%	90.0%
qwen3-14b	deepinfra/fp8	fp8	default	61.7%	23.3%	90.0%	46.7%	86.7%
qwen3.5-9b	deepinfra/bf16	bf16	default	59.2%	33.3%	76.7%	43.3%	83.3%
ministral-8b-2512	mistral/zdr	unknown	default	57.5%	50.0%	50.0%	53.3%	76.7%
qwen3-14b	nextbit/int4	int4	default	53.3%	20.0%	73.3%	40.0%	80.0%
qwen3.5-9b	darkbloom/fp4	fp4	default	52.5%	40.0%	66.7%	16.7%	86.7%
qwen3.5-9b	deepinfra/bf16	bf16	none	50.0%	36.7%	76.7%	33.3%	53.3%
ministral-3b-2512	mistral/zdr	unknown	default	33.3%	30.0%	40.0%	20.0%	43.3%
gemma-3-4b-it	deepinfra/bf16	bf16	default	20.9%	10.0%	46.7%	26.7%	0.0%

4 entrants have no score on at least one topic and are left out of this table. The topic tables below still show them.

AWS Quick Ten3 takes per question

Model	Host	Quant.	Thinking	Score	reports-read	arns	subnets	lifecycle	trust	query

Python Quick Ten3 takes per question

Model	Host	Quant.	Thinking	Score	is_leap	rle	chunks	round_half_up	parse_kv	merge_intervals

Rust Quick Ten3 takes per question

Model	Host	Quant.	Thinking	Score	is_palindrome	word_count	longest	parse_pair	compose	Fib

Tailwind Quick Ten3 takes per question

Model	Host	Quant.	Thinking	Score	box	row	t	wrap	b	card	g	panel