Model overview · DeepSeek
DeepSeek V3 (671B)
Generated from Architecture IR 0.1.0 · revision aec35a75d9
· report a correction
Overview
Fact sheet
- Layers
- 61
- Hidden size
- 7168
- Attention heads
- 128 · KV 128
- Vocab
- 129,280
- Context
- 160K
- Params
- 671B total / 37B active
Architecture Genome
One column per layer — top: attention, bottom: FFN/MoE.
Layer
attention (GQA/MLA) linear / recurrent attention MoE dense FFN
Evidence
Every visible number resolves to one of these claims. Inferred or unknown values are never shown as fact.
| Claim | Value | Status | Source |
|---|---|---|---|
facts.num_hidden_layers | 61 | • reported | hf_config · $.num_hidden_layers |
facts.hidden_size | 7168 | • reported | hf_config · $.hidden_size |
facts.num_attention_heads | 128 | • reported | hf_config · $.num_attention_heads |
facts.vocab_size | 129280 | • reported | hf_config · $.vocab_size |
facts.context_tokens | 163840 | • reported | hf_config · $.max_position_embeddings |
facts.total_params | 671000000000 | • reported | model_card · parameter count |
facts.num_key_value_heads | — | ? unknown | hf_config |
topology.attention_groups[0] | "61 MLA layers" | • reported | hf_config · layers config / architecture table |
topology.experts.routed_total | 256 | • reported | hf_config · $.n_routed_experts |
topology.experts.active_routed | 8 | • reported | hf_config · $.num_experts_per_tok |
topology.experts.shared | 1 | • reported | hf_config · $.n_shared_experts |
topology.position_encoding.type | "rope" | • reported | hf_config · rope settings |
facts.active_params | 37000000000 | = derived | derived · dense or activated experts + attention params |
Compare
Compare this model with any other catalog model (see catalog).