Architecture similarity

Each model is encoded as a feature vector — attention class, MoE layout, normalization, positional encoding, activation, dtype, plus log-scaled size and context. Cosine similarity between these vectors surfaces architectural siblings: models that made the same structural bets at different scales.

Method

Vectors are built from one-hot encoding of categorical choices (attention class, norm, position encoding, activation, dtype) plus normalized log-scale features (total parameters, context length, layers, hidden size). Similarity is cosine; it measures structural resemblance, not capability.

Similarity heatmap

01 Rows and columns are models, grouped by laboratory; darker means more similar. Hover for exact values.

Similarity values range from 0 (no shared categorical features) to 1 (identical profiles).

similarity > 0.93

Model A		Model B		similarity
DeepSeek-V3	DeepSeek	DeepSeek-V3 Base	DeepSeek	1.000
DeepSeek-V3	DeepSeek	DeepSeek-R1	DeepSeek	1.000
DeepSeek-V3	DeepSeek	DeepSeek-V3.1	DeepSeek	1.000
DeepSeek-V3	DeepSeek	DeepSeek-V3.2	DeepSeek	1.000
DeepSeek-V3 Base	DeepSeek	DeepSeek-R1	DeepSeek	1.000
DeepSeek-V3 Base	DeepSeek	DeepSeek-V3.1	DeepSeek	1.000
DeepSeek-V3 Base	DeepSeek	DeepSeek-V3.2	DeepSeek	1.000
DeepSeek-R1	DeepSeek	DeepSeek-V3.1	DeepSeek	1.000
DeepSeek-R1	DeepSeek	DeepSeek-V3.2	DeepSeek	1.000
DeepSeek-V3.1	DeepSeek	DeepSeek-V3.2	DeepSeek	1.000
GLM-5	GLM	GLM-5.1	GLM	1.000
DeepSeek-V4-Flash	DeepSeek	DeepSeek-V4-Flash Base	DeepSeek	1.000
Hunyuan 3 Preview	Hunyuan	Hunyuan 3 Preview Base	Hunyuan	1.000
Hunyuan 3 Preview	Hunyuan	Hunyuan 3	Hunyuan	1.000
Hunyuan 3 Preview Base	Hunyuan	Hunyuan 3	Hunyuan	1.000
MiniMax-M3	MiniMax	MiniMax-M3 MXFP8	MiniMax	1.000
Qwen2 72B	Qwen	Qwen2.5 72B	Qwen	1.000
Kimi K2 0905	Kimi	Kimi K2.5	Kimi	1.000
Kimi K2 0905	Kimi	Kimi K2.6	Kimi	1.000
Qwen3-Next 80B A3B	Qwen	Qwen3-Coder-Next	Qwen	1.000
Kimi K2.5	Kimi	Kimi K2.6	Kimi	1.000
Qwen3.5 9B	Qwen	Qwen3.5 9B Base	Qwen	1.000
Qwen3.6 27B	Qwen	Qwen3.8 27B	Qwen	1.000
MiniMax-Text-01	MiniMax	MiniMax-M1	MiniMax	1.000
GLM-5.2	GLM	GLM-5.3	GLM	1.000
MiniMax-M2	MiniMax	MiniMax-M2.5	MiniMax	1.000
GLM-5.2	GLM	Hunyuan 4 Preview	Hunyuan	1.000
GLM-5.3	GLM	Hunyuan 4 Preview	Hunyuan	1.000
Qwen3 30B A3B	Qwen	Hunyuan-TurboS 30B A3B	Hunyuan	1.000
Step-3.5-Flash	StepFun	Step-3.7-Flash	StepFun	1.000
MiniMax-M2	MiniMax	MiniMax-M2.7	MiniMax	1.000
MiniMax-M2.5	MiniMax	MiniMax-M2.7	MiniMax	1.000
Mistral 7B	Mistral	Hunyuan 7B	Hunyuan	1.000
GLM-4.5	GLM	GLM-4.6	GLM	1.000
Qwen3 8B	Qwen	Hunyuan 7B	Hunyuan	1.000
Kimi K2 Base	Kimi	Kimi K2 0905	Kimi	1.000
Kimi K2 Base	Kimi	Kimi K2.5	Kimi	1.000
Kimi K2 Base	Kimi	Kimi K2.6	Kimi	1.000
Mistral 7B	Mistral	Qwen3 8B	Qwen	1.000
Qwen2.5 7B	Qwen	Hunyuan 7B	Hunyuan	1.000

Similarity values range from 0 (no shared categorical features) to 1 (identical profiles).