nemotron-3.5-bolt:e4b

9,522 21 minutes ago

Nemotron 3.5 Bolt is NVIDIA’s compact vision-language model for local coding, reasoning, and tool use.

vision tools thinking audio e4b
... /
json
6447bf78b308 · 4.3kB
{
"architectures": [
"DFlashDraftModel"
],
"model_type": "qwen3",
"block_size": 8,
"dflash_config": {
"mask_token_id": 4,
"target_layer_ids": [
5,
11,
17,
23,
35
],
"use_swa": true,
"swa_window_size": 1024,
"causal": false
},
"hidden_size": 2560,
"num_hidden_layers": 5,
"num_attention_heads": 32,
"num_key_value_heads": 8,
"head_dim": 128,
"intermediate_size": 5120,
"hidden_act": "silu",
"rms_norm_eps": 1e-06,
"vocab_size": 262144,
"max_position_embeddings": 131072,
"initializer_range": 0.02,
"attention_bias": false,
"attention_dropout": 0.0,
"rope_theta": 1000000.0,
"rope_scaling": null,
"tie_word_embeddings": false,
"torch_dtype": "bfloat16",
"num_target_layers": 42,
"layer_types": [
"sliding_attention",
"sliding_attention",
"sliding_attention",
"sliding_attention",
"sliding_attention"
],
"sliding_window": 1024,
"has_own_lm_head": true,
"quantization_config": {
"quant_algo": "W4A16_NVFP4",
"kv_cache_quant_algo": null,
"group_size": 16,
"exclude_modules": [
"hidden_norm",
"*hidden_norm",
"layers.0.input_layernorm",
"*layers.0.input_layernorm",
"layers.0.post_attention_layernorm",
"*layers.0.post_attention_layernorm",
"layers.0.self_attn.k_norm",
"*layers.0.self_attn.k_norm",
"layers.0.self_attn.q_norm",
"*layers.0.self_attn.q_norm",
"layers.1.input_layernorm",
"*layers.1.input_layernorm",
"layers.1.post_attention_layernorm",
"*layers.1.post_attention_layernorm",
"layers.1.self_attn.k_norm",
"*layers.1.self_attn.k_norm",
"layers.1.self_attn.q_norm",
"*layers.1.self_attn.q_norm",
"layers.2.input_layernorm",
"*layers.2.input_layernorm",
"layers.2.post_attention_layernorm",
"*layers.2.post_attention_layernorm",
"layers.2.self_attn.k_norm",
"*layers.2.self_attn.k_norm",
"layers.2.self_attn.q_norm",
"*layers.2.self_attn.q_norm",
"layers.3.input_layernorm",
"*layers.3.input_layernorm",
"layers.3.post_attention_layernorm",
"*layers.3.post_attention_layernorm",
"layers.3.self_attn.k_norm",
"*layers.3.self_attn.k_norm",
"layers.3.self_attn.q_norm",
"*layers.3.self_attn.q_norm",
"layers.4.input_layernorm",
"*layers.4.input_layernorm",
"layers.4.post_attention_layernorm",
"*layers.4.post_attention_layernorm",
"layers.4.self_attn.k_norm",
"*layers.4.self_attn.k_norm",
"layers.4.self_attn.q_norm",
"*layers.4.self_attn.q_norm",
"norm",
"*norm"
],
"quant_method": "modelopt_fp4",
"ignore": [
"hidden_norm",
"*hidden_norm",
"layers.0.input_layernorm",
"*layers.0.input_layernorm",
"layers.0.post_attention_layernorm",
"*layers.0.post_attention_layernorm",
"layers.0.self_attn.k_norm",
"*layers.0.self_attn.k_norm",
"layers.0.self_attn.q_norm",
"*layers.0.self_attn.q_norm",
"layers.1.input_layernorm",
"*layers.1.input_layernorm",
"layers.1.post_attention_layernorm",
"*layers.1.post_attention_layernorm",
"layers.1.self_attn.k_norm",
"*layers.1.self_attn.k_norm",
"layers.1.self_attn.q_norm",
"*layers.1.self_attn.q_norm",
"layers.2.input_layernorm",
"*layers.2.input_layernorm",
"layers.2.post_attention_layernorm",
"*layers.2.post_attention_layernorm",
"layers.2.self_attn.k_norm",
"*layers.2.self_attn.k_norm",
"layers.2.self_attn.q_norm",
"*layers.2.self_attn.q_norm",
"layers.3.input_layernorm",
"*layers.3.input_layernorm",
"layers.3.post_attention_layernorm",
"*layers.3.post_attention_layernorm",
"layers.3.self_attn.k_norm",
"*layers.3.self_attn.k_norm",
"layers.3.self_attn.q_norm",
"*layers.3.self_attn.q_norm",
"layers.4.input_layernorm",
"*layers.4.input_layernorm",
"layers.4.post_attention_layernorm",
"*layers.4.post_attention_layernorm",
"layers.4.self_attn.k_norm",
"*layers.4.self_attn.k_norm",
"layers.4.self_attn.q_norm",
"*layers.4.self_attn.q_norm",
"norm",
"*norm"
]
}
}