tinker-nomics

Models

Model specifications for all Tinker-supported models.

Model	Org	Arch	Total	Active	Layers	d_model	Attention	Experts
Qwen3.5-397B-A17B	qwen	moe	397B	17B	—	—	GQA	—
Qwen3.5-35B-A3B	qwen	moe	35B	3B	—	—	GQA	—
Qwen3.5-27B	qwen	dense	27B	27B	—	—	GQA	—
Qwen3.5-4B	qwen	dense	4B	4B	—	—	GQA	—
Qwen3-VL-235B-A22B-Instruct	qwen	moe	235B	22B	94	4096	GQA	128/8
Qwen3-VL-30B-A3B-Instruct	qwen	moe	30.5B	3.3B	48	2048	GQA	128/8
Qwen3-235B-A22B-Instruct-2507	qwen	moe	235B	22B	94	4096	GQA	128/8
Qwen3-30B-A3B-Instruct-2507	qwen	moe	30.5B	3.3B	48	2048	GQA	128/8
Qwen3-30B-A3B	qwen	moe	30.5B	3.3B	48	2048	GQA	128/8
Qwen3-30B-A3B-Base	qwen	moe	30.5B	3.3B	48	2048	GQA	128/8
Qwen3-32B	qwen	dense	32B	32B	64	5120	GQA	—
Qwen3-8B	qwen	dense	8B	8B	36	4096	GQA	—
Qwen3-8B-Base	qwen	dense	8B	8B	36	4096	GQA	—
Qwen3-4B-Instruct-2507	qwen	dense	4B	4B	36	2560	GQA	—
GPT-OSS-120B	openai	moe	117B	5.1B	—	—	GQA	—
GPT-OSS-20B	openai	moe	21B	3.6B	—	—	GQA	—
DeepSeek-V3.1	deepseek	moe	671B	37B	61	7168	MLA	256/8
DeepSeek-V3.1-Base	deepseek	moe	671B	37B	61	7168	MLA	256/8
Llama-3.1-70B	meta	dense	70B	70B	80	8192	GQA	—
Llama-3.3-70B-Instruct	meta	dense	70B	70B	80	8192	GQA	—
Llama-3.1-8B	meta	dense	8B	8B	32	4096	GQA	—
Llama-3.1-8B-Instruct	meta	dense	8B	8B	32	4096	GQA	—
Llama-3.2-3B	meta	dense	3B	3B	28	3072	GQA	—
Llama-3.2-1B	meta	dense	1B	1B	16	2048	GQA	—
Kimi-K2-Thinking	moonshot	moe	1000B	32B	—	—	MLA	—
Kimi-K2.5	moonshot	moe	1000B	32B	—	—	MLA	—