Whisper large-v1
dense 1.55B
Fields
id
models/whisper-large-v1Family
whisper
Vendor
OpenAI
Release year
2022
Architecture
dense
Params (B)
1.55
Active params (B)
1.55
Hidden size
1,280
Layers
32
KV heads
20
Head dim
64
GQA ratio
1
Max context
1,500
Context scaling method
native
Activation
gelu
Attention variant
MHA, encoder-decoder (cross-attention)
KV cache (bytes/token)
163,840
FLOPs per active token
3,100,000,000
Open weights
yes
License
apache-2.0
notes
ENCODER-DECODER. DRAFT RECORD -- config.json could not be fetched from HuggingFace (500 error). Architecture is inferred from large-v2 (same d_model, layers, heads) but num_mel_bins is assumed to be 80 (not verified). ENCODER: 32 layers, d_model 1280, 20 heads, encoder_ffn_dim 5120, gelu, max_source_positions 1500. DECODER: 32 layers, 20 heads, decoder_ffn_dim 5120, max_target_positions 448. KV CACHE: decoder self-attention 32 * 20 * 64 * 2 * 2 = 163,840 B/token = 160 KiB/token. Cross-attention over 1500 encoder frames: 32 * 20 * 64 * 1500 * 2 * 2 = 245,760,000 B = 234 MiB per in-flight request. STREAMING: not natively supported (offline/batch model).
task
speech-to-text
Notes
ENCODER-DECODER. DRAFT RECORD -- config.json could not be fetched from HuggingFace (500 error). Architecture is inferred from large-v2 (same d_model, layers, heads) but num_mel_bins is assumed to be 80 (not verified). ENCODER: 32 layers, d_model 1280, 20 heads, encoder_ffn_dim 5120, gelu, max_source_positions 1500. DECODER: 32 layers, 20 heads, decoder_ffn_dim 5120, max_target_positions 448. KV CACHE: decoder self-attention 32 * 20 * 64 * 2 * 2 = 163,840 B/token = 160 KiB/token. Cross-attention over 1500 encoder frames: 32 * 20 * 64 * 1500 * 2 * 2 = 245,760,000 B = 234 MiB per in-flight request. STREAMING: not natively supported (offline/batch model).
Cross-references
This record references no other records — it is a leaf. It is cited by 0.