Tested models · EshAlora

The models we tested

Everything that has ever gone through one of the measurements, from two and a half billion parameters to almost three trillion. Sorted by size.

For MoE models there are two numbers: how many parameters the model has in total and how many actually work on a single token. Those are two different things, and the first one alone would distort the table.

Huge

two trillion parameters and up

model	maker	size	run through
Kimi K3	Moonshot	2.8 T	API
Qwen 3.8 Max	Alibaba	2.4 T / 95 B active	API
Fable 5	Anthropic	huge\*	web
Fable 5.1	Anthropic	huge\*	API
Gemini 3.1 Pro	Google	huge\*	API
GPT-6 Astra	OpenAI	huge\*	API
GPT-6 Astra Pro	OpenAI	huge\*	API

Large

one to two trillion

model	maker	size	run through
DeepSeek V4 Pro	DeepSeek	1.6 T / 49 B active	API
LongCat 2.0	Meituan	1.6 T / 48 B active	API
MiMo 2.5 Pro	Xiaomi	1 T / 42 B active	API
GPT-5.6 Sol	OpenAI	large\*	API
Grok 4.6	xAI	large\*	API
Muse Spark 1.3	Meta	large\*	API
Opus 4.6	Anthropic	large\*	web
Opus 4.7	Anthropic	large\*	web
Opus 4.8	Anthropic	large\*	web
Opus 5	Anthropic	large\*	API

Mid-sized

three hundred billion to a trillion

model	maker	size	run through
Inkling	Thinking Machines	975 B / 41 B active	API
Hunyuan 4	Tencent	770 B / 49 B active	API
GLM 5.3	Zhipu	753 B / 39 B active	API
DeepSeek V3.2	DeepSeek	685 B / 37 B active	API
Mistral Large 3 2512	Mistral	675 B / 41 B active	API
Nemotron 3 Ultra	NVIDIA	550 B / 55 B active	API
MiniMax M3	MiniMax	427 B / 26 B active	API
Ernie 4.5 VL	Baidu	424 B / 47 B active	API
Llama 4 Maverick	Meta	400 B / 17 B active	API
Trinity Large	Arcee	398 B / 13 B active	API
Qwen 3.5 397B-A17B	Alibaba	397 B / 17 B active	API
Aion 2.0	AionLabs	mid-sized\*	API
Aion 3.0	AionLabs	mid-sized\*	API
Aion 3.0 Mini	AionLabs	mid-sized\*	API
GPT-5.6 Terra	OpenAI	mid-sized\*	API
GPT-5.6 Terra Pro	OpenAI	mid-sized\*	API
Nova Premier	Amazon	mid-sized\*	API
Opus 3	Anthropic	mid-sized\*	web
Seed 2.1 Turbo	ByteDance	mid-sized\*	API
Sonnet 4.6	Anthropic	mid-sized\*	web
Sonnet 5	Anthropic	mid-sized\*	API

Small

fifty to three hundred billion

model	maker	size	run through
Step 3.7 Flash	StepFun	196 B / 11 B active	API
Mistral Medium 3.5	Mistral	128 B	API
Ling 3.0 Flash	Ant Group	124 B / 5.1 B active	API
Qwen 3.5 122B-A10B	Alibaba	122 B / 10 B active	API
Command A	Cohere	111 B	API
Euryale L3.3 70B	Sao10K	70 B	API
Hermes 4 70B	Nous Research	70 B	API
Llama 3.1 70B	Meta	70 B	API
Llama 3.3 70B	Meta	70 B	API
Sellma	Seznam	70 B	web
GPT-5.6 Luna	OpenAI	small\*	API
GPT-5.6 Luna Pro	OpenAI	small\*	API
Haiku 4.5	Anthropic	small\*	API
MiniMax M2 her	MiniMax	small\*	API
Solar Pro 4	Upstage	small\*	API

Tiny

fifteen to fifty billion

model	maker	size	run through
Qwen 3.5 35B-A3B	Alibaba	35 B / 3 B active	API
Gemma 4 31B QAT	Google	31 B	local
Muse Glimmer 30B	Meta	30 B	local
Qwen 3.5 27B	Alibaba	27 B	API
Qwen 3.8 27B	Alibaba	27 B	API
Gemma 4 26B A4B	Google	26 B / 4 B active	API
Cydonia 24B v4.1	TheDrummer	24 B	API
Dolphin Mistral 24B Venice	Cognitive Computations	24 B	API
Mistral Small 24B 2501	Mistral	24 B	API
Mistral Small 3.2 24B	Mistral	24 B	API
Reka Flash 3	Reka	21 B	API
Mercury 2.5	Inception	tiny\*	API

Miniature

under fifteen billion

model	maker	size	run through
MythoMax 13B	Gryphe	13 B	API
Mistral Nemo 12B	Mistral	12 B	API
UnslopNemo 12B	TheDrummer	12 B	API
Qwen 3.5 9B	Alibaba	9 B	API
Aion RP Llama 3.1 8B	AionLabs	8 B	API
Granite 4.2 8B	IBM	8 B	API
Llama 3.1 8B	Meta	8 B	API
Gemma 3 4B	Google	4 B	API
LFM 2.5 2.6B	Liquid AI	2.6 B	API
Gemma 4 E2B	Google	2 B	local
Weaver	Mancer	miniature\*	API

\* An asterisk means the maker does not publish the size. The band is not an estimate of a parameter count. It is a placement by how the maker itself ranks the model in its own line-up, and against the models in this table where the number is known. The figures that circulate about closed models contradict each other and cannot be verified, so they are not here.

This is not a list of models that completed every measurement. It is a list of those that went through at least one test. Some runs ended in an error, a filter or a collapse, and they are here all the same.