flint:latest

1 13 hours ago

Flint is a 39.8 billion-parameter mixture-of-experts model from Recursive, built for coding agents.

tools thinking 40b
... /
json
3683212f2558 ยท 3.1kB
{
"architectures": [
"FlintForCausalLM"
],
"attention_bias": false,
"attention_dropout": 0.0,
"attention_sink": true,
"bos_token_id": 150000,
"decoder_sparse_step": 1,
"dtype": "bfloat16",
"eos_token_id": [
150001,
150003,
150005
],
"head_dim": 128,
"hidden_act": "silu",
"hidden_size": 2048,
"intermediate_size": 6144,
"kv_group_size": 64,
"kv_key_bits": 8,
"kv_value_bits": 8,
"layer_types": [
"full_attention",
"full_attention",
"sliding_attention",
"sliding_attention",
"sliding_attention",
"full_attention",
"sliding_attention",
"sliding_attention",
"sliding_attention",
"full_attention",
"sliding_attention",
"sliding_attention",
"sliding_attention",
"full_attention",
"sliding_attention",
"sliding_attention",
"sliding_attention",
"full_attention",
"sliding_attention",
"sliding_attention",
"sliding_attention",
"full_attention",
"sliding_attention",
"sliding_attention",
"sliding_attention",
"full_attention",
"sliding_attention",
"sliding_attention",
"sliding_attention",
"full_attention",
"sliding_attention",
"sliding_attention",
"sliding_attention",
"full_attention",
"sliding_attention",
"sliding_attention",
"sliding_attention",
"full_attention",
"sliding_attention",
"sliding_attention",
"sliding_attention",
"full_attention",
"sliding_attention",
"sliding_attention",
"sliding_attention",
"full_attention",
"sliding_attention",
"sliding_attention",
"full_attention"
],
"max_position_embeddings": 262144,
"mlp_only_layers": [
0,
1
],
"mlx_export": {
"bits": 4,
"embedding_bits": 8,
"full_attention_layers": [
0,
1,
5,
9,
13,
17,
21,
25,
29,
33,
37,
41,
45,
48
],
"routed_expert_bits": 4,
"scale_rule": "mse"
},
"model_type": "flint",
"moe_intermediate_size": 512,
"no_rope_layer_indices": [
0,
1,
5,
9,
13,
17,
21,
25,
29,
33,
37,
41,
45,
48
],
"norm_topk_prob": true,
"num_attention_heads": 32,
"num_experts": 256,
"num_experts_per_tok": 6,
"num_hidden_layers": 49,
"num_key_value_heads": 4,
"num_local_experts": 256,
"output_router_logits": false,
"quantization": {
"bits": 4,
"group_size": 64,
"mode": "affine"
},
"rms_norm_eps": 1e-06,
"rope_parameters": {
"rope_theta": 10000.0,
"rope_type": "default"
},
"rope_theta": 10000.0,
"router_aux_loss_coef": 0.0,
"router_expert_bias": false,
"scoring_func": "sigmoid",
"shared_expert_intermediate_size": 1024,
"sliding_window": 513,
"sliding_window_keys": 513,
"softmax_type": "learnable",
"softmax_type_source": "model.attention_pattern.softmax_type",
"tie_word_embeddings": false,
"use_cache": true,
"use_sliding_window": true,
"vocab_size": 150144,
"router_score_function": "sigmoid"
}