AntLing/ Ling-3.0-tiny:q8_0

201 2 weeks ago

Ling-3.0-tiny, a lightweight hybrid reasoning MoE model with 7.9B total parameters and only 1.3B activated parameters per token. It is designed to deliver strong reasoning and agentic capabilities at low inference cost, making advanced model capabilities.

thinking
... /
model
28d3a3e1b849 · 8.4GB
    Metadata
  • general.architecture
    bailingmoe3
  • general.file_type
    Q8_0
  • bailingmoe3.attention.head_count
    16
  • bailingmoe3.attention.head_count_kv
    [0, 0, 0, 1, 0, ...]
  • bailingmoe3.attention.key_length
    576
  • bailingmoe3.attention.key_length_mla
    192
  • bailingmoe3.attention.kv_lora_rank
    512
  • bailingmoe3.attention.layer_norm_rms_epsilon
    1e-06
  • bailingmoe3.attention.q_lora_rank
    256
  • bailingmoe3.attention.value_length
    128
  • bailingmoe3.attention.value_length_mla
    128
  • bailingmoe3.block_count
    24
  • bailingmoe3.context_length
    131072
  • bailingmoe3.embedding_length
    1536
  • bailingmoe3.expert_count
    128
  • bailingmoe3.expert_feed_forward_length
    512
  • bailingmoe3.expert_gating_func
    2
  • bailingmoe3.expert_group_count
    8
  • bailingmoe3.expert_group_used_count
    4
  • bailingmoe3.expert_shared_count
    1
  • bailingmoe3.expert_shared_feed_forward_length
    512
  • bailingmoe3.expert_used_count
    8
  • bailingmoe3.expert_weights_norm
    true
  • bailingmoe3.expert_weights_scale
    2.5
  • bailingmoe3.feed_forward_length
    4608
  • bailingmoe3.kda.gate_lower_bound
    -5
  • bailingmoe3.kda.head_dim
    128
  • bailingmoe3.kda.safe_gate
    true
  • bailingmoe3.leading_dense_block_count
    1
  • bailingmoe3.rope.dimension_count
    64
  • bailingmoe3.rope.freq_base
    6e+06
  • bailingmoe3.ssm.conv_kernel
    4
  • bailingmoe3.vocab_size
    157184
  • tokenizer.ggml.add_bos_token
    false
  • tokenizer.ggml.add_eos_token
    false
  • tokenizer.ggml.bos_token_id
    156891
  • tokenizer.ggml.eos_token_id
    156895
  • tokenizer.ggml.merges
    [Ġ Ġ, Ġ t, i n, Ġ a, h e, ...]
  • tokenizer.ggml.model
    gpt2
  • tokenizer.ggml.padding_token_id
    156892
  • tokenizer.ggml.pre
    bailingmoe2
  • tokenizer.ggml.token_type
    [1, 1, 1, 1, 1, ...]
  • tokenizer.ggml.tokens
    [!, ", #, $, %, ...]
    • Tensor
  • token_embd.weight
    Q8_0
    [1536, 157184]
    • blk.0
  • blk.0.attn_k.weight
    Q8_0
    [1536, 2048]
  • blk.0.attn_norm.weight
    F32
    [1536]
  • blk.0.attn_output.weight
    Q8_0
    [2048, 1536]
  • blk.0.attn_q.weight
    Q8_0
    [1536, 2048]
  • blk.0.attn_v.weight
    Q8_0
    [1536, 2048]
  • blk.0.ffn_down.weight
    Q8_0
    [4608, 1536]
  • blk.0.ffn_gate.weight
    Q8_0
    [1536, 4608]
  • blk.0.ffn_norm.weight
    F32
    [1536]
  • blk.0.ffn_up.weight
    Q8_0
    [1536, 4608]
  • blk.0.ssm_a
    F32
    [1, 16]
  • blk.0.ssm_beta.weight
    Q8_0
    [1536, 16]
  • blk.0.ssm_conv1d_k.weight
    F32
    [4, 1, 2048]
  • blk.0.ssm_conv1d_q.weight
    F32
    [4, 1, 2048]
  • blk.0.ssm_conv1d_v.weight
    F32
    [4, 1, 2048]
  • blk.0.ssm_dt.bias
    F32
    [2048]
  • blk.0.ssm_f_a.weight
    Q8_0
    [1536, 2048]
  • blk.0.ssm_g_a.weight
    Q8_0
    [1536, 2048]
  • blk.0.ssm_norm.weight
    F32
    [128]
    • blk.1
  • blk.1.attn_k.weight
    Q8_0
    [1536, 2048]
  • blk.1.attn_norm.weight
    F32
    [1536]
  • blk.1.attn_output.weight
    Q8_0
    [2048, 1536]
  • blk.1.attn_q.weight
    Q8_0
    [1536, 2048]
  • blk.1.attn_v.weight
    Q8_0
    [1536, 2048]
  • blk.1.exp_probs_b.bias
    F32
    [128]
  • blk.1.ffn_down_exps.weight
    Q8_0
    [512, 1536, 128]
  • blk.1.ffn_down_shexp.weight
    Q8_0
    [512, 1536]
  • blk.1.ffn_gate_exps.weight
    Q8_0
    [1536, 512, 128]
  • blk.1.ffn_gate_inp.weight
    F32
    [1536, 128]
  • blk.1.ffn_gate_shexp.weight
    Q8_0
    [1536, 512]
  • blk.1.ffn_norm.weight
    F32
    [1536]
  • blk.1.ffn_up_exps.weight
    Q8_0
    [1536, 512, 128]
  • blk.1.ffn_up_shexp.weight
    Q8_0
    [1536, 512]
  • blk.1.ssm_a
    F32
    [1, 16]
  • blk.1.ssm_beta.weight
    Q8_0
    [1536, 16]
  • blk.1.ssm_conv1d_k.weight
    F32
    [4, 1, 2048]
  • blk.1.ssm_conv1d_q.weight
    F32
    [4, 1, 2048]
  • blk.1.ssm_conv1d_v.weight
    F32
    [4, 1, 2048]
  • blk.1.ssm_dt.bias
    F32
    [2048]
  • blk.1.ssm_f_a.weight
    Q8_0
    [1536, 2048]
  • blk.1.ssm_g_a.weight
    Q8_0
    [1536, 2048]
  • blk.1.ssm_norm.weight
    F32
    [128]
    • blk.2
  • blk.2.attn_k.weight
    Q8_0
    [1536, 2048]
  • blk.2.attn_norm.weight
    F32
    [1536]
  • blk.2.attn_output.weight
    Q8_0
    [2048, 1536]
  • blk.2.attn_q.weight
    Q8_0
    [1536, 2048]
  • blk.2.attn_v.weight
    Q8_0
    [1536, 2048]
  • blk.2.exp_probs_b.bias
    F32
    [128]
  • blk.2.ffn_down_exps.weight
    Q8_0
    [512, 1536, 128]
  • blk.2.ffn_down_shexp.weight
    Q8_0
    [512, 1536]
  • blk.2.ffn_gate_exps.weight
    Q8_0
    [1536, 512, 128]
  • blk.2.ffn_gate_inp.weight
    F32
    [1536, 128]
  • blk.2.ffn_gate_shexp.weight
    Q8_0
    [1536, 512]
  • blk.2.ffn_norm.weight
    F32
    [1536]
  • blk.2.ffn_up_exps.weight
    Q8_0
    [1536, 512, 128]
  • blk.2.ffn_up_shexp.weight
    Q8_0
    [1536, 512]
  • blk.2.ssm_a
    F32
    [1, 16]
  • blk.2.ssm_beta.weight
    Q8_0
    [1536, 16]
  • blk.2.ssm_conv1d_k.weight
    F32
    [4, 1, 2048]
  • blk.2.ssm_conv1d_q.weight
    F32
    [4, 1, 2048]
  • blk.2.ssm_conv1d_v.weight
    F32
    [4, 1, 2048]
  • blk.2.ssm_dt.bias
    F32
    [2048]
  • blk.2.ssm_f_a.weight
    Q8_0
    [1536, 2048]
  • blk.2.ssm_g_a.weight
    Q8_0
    [1536, 2048]
  • blk.2.ssm_norm.weight
    F32
    [128]
    • blk.3
  • blk.3.attn_gate.weight
    Q8_0
    [1536, 16]
  • blk.3.attn_k_b.weight
    Q8_0
    [128, 512, 16]
  • blk.3.attn_kv_a_mqa.weight
    Q8_0
    [1536, 576]
  • blk.3.attn_kv_a_norm.weight
    F32
    [512]
  • blk.3.attn_norm.weight
    F32
    [1536]
  • blk.3.attn_output.weight
    Q8_0
    [2048, 1536]
  • blk.3.attn_q_a.weight
    Q8_0
    [1536, 256]
  • blk.3.attn_q_a_norm.weight
    F32
    [256]
  • blk.3.attn_q_b.weight
    Q8_0
    [256, 3072]
  • blk.3.attn_v_b.weight
    Q8_0
    [512, 128, 16]
  • blk.3.exp_probs_b.bias
    F32
    [128]
  • blk.3.ffn_down_exps.weight
    Q8_0
    [512, 1536, 128]
  • blk.3.ffn_down_shexp.weight
    Q8_0
    [512, 1536]
  • blk.3.ffn_gate_exps.weight
    Q8_0
    [1536, 512, 128]
  • blk.3.ffn_gate_inp.weight
    F32
    [1536, 128]
  • blk.3.ffn_gate_shexp.weight
    Q8_0
    [1536, 512]
  • blk.3.ffn_norm.weight
    F32
    [1536]
  • blk.3.ffn_up_exps.weight
    Q8_0
    [1536, 512, 128]
  • blk.3.ffn_up_shexp.weight
    Q8_0
    [1536, 512]
    • blk.4
  • blk.4.attn_k.weight
    Q8_0
    [1536, 2048]
  • blk.4.attn_norm.weight
    F32
    [1536]
  • blk.4.attn_output.weight
    Q8_0
    [2048, 1536]
  • blk.4.attn_q.weight
    Q8_0
    [1536, 2048]
  • blk.4.attn_v.weight
    Q8_0
    [1536, 2048]
  • blk.4.exp_probs_b.bias
    F32
    [128]
  • blk.4.ffn_down_exps.weight
    Q8_0
    [512, 1536, 128]
  • blk.4.ffn_down_shexp.weight
    Q8_0
    [512, 1536]
  • blk.4.ffn_gate_exps.weight
    Q8_0
    [1536, 512, 128]
  • blk.4.ffn_gate_inp.weight
    F32
    [1536, 128]
  • blk.4.ffn_gate_shexp.weight
    Q8_0
    [1536, 512]
  • blk.4.ffn_norm.weight
    F32
    [1536]
  • blk.4.ffn_up_exps.weight
    Q8_0
    [1536, 512, 128]
  • blk.4.ffn_up_shexp.weight
    Q8_0
    [1536, 512]
  • blk.4.ssm_a
    F32
    [1, 16]
  • blk.4.ssm_beta.weight
    Q8_0
    [1536, 16]
  • blk.4.ssm_conv1d_k.weight
    F32
    [4, 1, 2048]
  • blk.4.ssm_conv1d_q.weight
    F32
    [4, 1, 2048]
  • blk.4.ssm_conv1d_v.weight
    F32
    [4, 1, 2048]
  • blk.4.ssm_dt.bias
    F32
    [2048]
  • blk.4.ssm_f_a.weight
    Q8_0
    [1536, 2048]
  • blk.4.ssm_g_a.weight
    Q8_0
    [1536, 2048]
  • blk.4.ssm_norm.weight
    F32
    [128]
    • blk.5
  • blk.5.attn_k.weight
    Q8_0
    [1536, 2048]
  • blk.5.attn_norm.weight
    F32
    [1536]
  • blk.5.attn_output.weight
    Q8_0
    [2048, 1536]
  • blk.5.attn_q.weight
    Q8_0
    [1536, 2048]
  • blk.5.attn_v.weight
    Q8_0
    [1536, 2048]
  • blk.5.exp_probs_b.bias
    F32
    [128]
  • blk.5.ffn_down_exps.weight
    Q8_0
    [512, 1536, 128]
  • blk.5.ffn_down_shexp.weight
    Q8_0
    [512, 1536]
  • blk.5.ffn_gate_exps.weight
    Q8_0
    [1536, 512, 128]
  • blk.5.ffn_gate_inp.weight
    F32
    [1536, 128]
  • blk.5.ffn_gate_shexp.weight
    Q8_0
    [1536, 512]
  • blk.5.ffn_norm.weight
    F32
    [1536]
  • blk.5.ffn_up_exps.weight
    Q8_0
    [1536, 512, 128]
  • blk.5.ffn_up_shexp.weight
    Q8_0
    [1536, 512]
  • blk.5.ssm_a
    F32
    [1, 16]
  • blk.5.ssm_beta.weight
    Q8_0
    [1536, 16]
  • blk.5.ssm_conv1d_k.weight
    F32
    [4, 1, 2048]
  • blk.5.ssm_conv1d_q.weight
    F32
    [4, 1, 2048]
  • blk.5.ssm_conv1d_v.weight
    F32
    [4, 1, 2048]
  • blk.5.ssm_dt.bias
    F32
    [2048]
  • blk.5.ssm_f_a.weight
    Q8_0
    [1536, 2048]
  • blk.5.ssm_g_a.weight
    Q8_0
    [1536, 2048]
  • blk.5.ssm_norm.weight
    F32
    [128]
    • blk.6
  • blk.6.attn_k.weight
    Q8_0
    [1536, 2048]
  • blk.6.attn_norm.weight
    F32
    [1536]
  • blk.6.attn_output.weight
    Q8_0
    [2048, 1536]
  • blk.6.attn_q.weight
    Q8_0
    [1536, 2048]
  • blk.6.attn_v.weight
    Q8_0
    [1536, 2048]
  • blk.6.exp_probs_b.bias
    F32
    [128]
  • blk.6.ffn_down_exps.weight
    Q8_0
    [512, 1536, 128]
  • blk.6.ffn_down_shexp.weight
    Q8_0
    [512, 1536]
  • blk.6.ffn_gate_exps.weight
    Q8_0
    [1536, 512, 128]
  • blk.6.ffn_gate_inp.weight
    F32
    [1536, 128]
  • blk.6.ffn_gate_shexp.weight
    Q8_0
    [1536, 512]
  • blk.6.ffn_norm.weight
    F32
    [1536]
  • blk.6.ffn_up_exps.weight
    Q8_0
    [1536, 512, 128]
  • blk.6.ffn_up_shexp.weight
    Q8_0
    [1536, 512]
  • blk.6.ssm_a
    F32
    [1, 16]
  • blk.6.ssm_beta.weight
    Q8_0
    [1536, 16]
  • blk.6.ssm_conv1d_k.weight
    F32
    [4, 1, 2048]
  • blk.6.ssm_conv1d_q.weight
    F32
    [4, 1, 2048]
  • blk.6.ssm_conv1d_v.weight
    F32
    [4, 1, 2048]
  • blk.6.ssm_dt.bias
    F32
    [2048]
  • blk.6.ssm_f_a.weight
    Q8_0
    [1536, 2048]
  • blk.6.ssm_g_a.weight
    Q8_0
    [1536, 2048]
  • blk.6.ssm_norm.weight
    F32
    [128]
    • blk.7
  • blk.7.attn_gate.weight
    Q8_0
    [1536, 16]
  • blk.7.attn_k_b.weight
    Q8_0
    [128, 512, 16]
  • blk.7.attn_kv_a_mqa.weight
    Q8_0
    [1536, 576]
  • blk.7.attn_kv_a_norm.weight
    F32
    [512]
  • blk.7.attn_norm.weight
    F32
    [1536]
  • blk.7.attn_output.weight
    Q8_0
    [2048, 1536]
  • blk.7.attn_q_a.weight
    Q8_0
    [1536, 256]
  • blk.7.attn_q_a_norm.weight
    F32
    [256]
  • blk.7.attn_q_b.weight
    Q8_0
    [256, 3072]
  • blk.7.attn_v_b.weight
    Q8_0
    [512, 128, 16]
  • blk.7.exp_probs_b.bias
    F32
    [128]
  • blk.7.ffn_down_exps.weight
    Q8_0
    [512, 1536, 128]
  • blk.7.ffn_down_shexp.weight
    Q8_0
    [512, 1536]
  • blk.7.ffn_gate_exps.weight
    Q8_0
    [1536, 512, 128]
  • blk.7.ffn_gate_inp.weight
    F32
    [1536, 128]
  • blk.7.ffn_gate_shexp.weight
    Q8_0
    [1536, 512]
  • blk.7.ffn_norm.weight
    F32
    [1536]
  • blk.7.ffn_up_exps.weight
    Q8_0
    [1536, 512, 128]
  • blk.7.ffn_up_shexp.weight
    Q8_0
    [1536, 512]
    • blk.8
  • blk.8.attn_k.weight
    Q8_0
    [1536, 2048]
  • blk.8.attn_norm.weight
    F32
    [1536]
  • blk.8.attn_output.weight
    Q8_0
    [2048, 1536]
  • blk.8.attn_q.weight
    Q8_0
    [1536, 2048]
  • blk.8.attn_v.weight
    Q8_0
    [1536, 2048]
  • blk.8.exp_probs_b.bias
    F32
    [128]
  • blk.8.ffn_down_exps.weight
    Q8_0
    [512, 1536, 128]
  • blk.8.ffn_down_shexp.weight
    Q8_0
    [512, 1536]
  • blk.8.ffn_gate_exps.weight
    Q8_0
    [1536, 512, 128]
  • blk.8.ffn_gate_inp.weight
    F32
    [1536, 128]
  • blk.8.ffn_gate_shexp.weight
    Q8_0
    [1536, 512]
  • blk.8.ffn_norm.weight
    F32
    [1536]
  • blk.8.ffn_up_exps.weight
    Q8_0
    [1536, 512, 128]
  • blk.8.ffn_up_shexp.weight
    Q8_0
    [1536, 512]
  • blk.8.ssm_a
    F32
    [1, 16]
  • blk.8.ssm_beta.weight
    Q8_0
    [1536, 16]
  • blk.8.ssm_conv1d_k.weight
    F32
    [4, 1, 2048]
  • blk.8.ssm_conv1d_q.weight
    F32
    [4, 1, 2048]
  • blk.8.ssm_conv1d_v.weight
    F32
    [4, 1, 2048]
  • blk.8.ssm_dt.bias
    F32
    [2048]
  • blk.8.ssm_f_a.weight
    Q8_0
    [1536, 2048]
  • blk.8.ssm_g_a.weight
    Q8_0
    [1536, 2048]
  • blk.8.ssm_norm.weight
    F32
    [128]
    • blk.9
  • blk.9.attn_k.weight
    Q8_0
    [1536, 2048]
  • blk.9.attn_norm.weight
    F32
    [1536]
  • blk.9.attn_output.weight
    Q8_0
    [2048, 1536]
  • blk.9.attn_q.weight
    Q8_0
    [1536, 2048]
  • blk.9.attn_v.weight
    Q8_0
    [1536, 2048]
  • blk.9.exp_probs_b.bias
    F32
    [128]
  • blk.9.ffn_down_exps.weight
    Q8_0
    [512, 1536, 128]
  • blk.9.ffn_down_shexp.weight
    Q8_0
    [512, 1536]
  • blk.9.ffn_gate_exps.weight
    Q8_0
    [1536, 512, 128]
  • blk.9.ffn_gate_inp.weight
    F32
    [1536, 128]
  • blk.9.ffn_gate_shexp.weight
    Q8_0
    [1536, 512]
  • blk.9.ffn_norm.weight
    F32
    [1536]
  • blk.9.ffn_up_exps.weight
    Q8_0
    [1536, 512, 128]
  • blk.9.ffn_up_shexp.weight
    Q8_0
    [1536, 512]
  • blk.9.ssm_a
    F32
    [1, 16]
  • blk.9.ssm_beta.weight
    Q8_0
    [1536, 16]
  • blk.9.ssm_conv1d_k.weight
    F32
    [4, 1, 2048]
  • blk.9.ssm_conv1d_q.weight
    F32
    [4, 1, 2048]
  • blk.9.ssm_conv1d_v.weight
    F32
    [4, 1, 2048]
  • blk.9.ssm_dt.bias
    F32
    [2048]
  • blk.9.ssm_f_a.weight
    Q8_0
    [1536, 2048]
  • blk.9.ssm_g_a.weight
    Q8_0
    [1536, 2048]
  • blk.9.ssm_norm.weight
    F32
    [128]
    • blk.10
  • blk.10.attn_k.weight
    Q8_0
    [1536, 2048]
  • blk.10.attn_norm.weight
    F32
    [1536]
  • blk.10.attn_output.weight
    Q8_0
    [2048, 1536]
  • blk.10.attn_q.weight
    Q8_0
    [1536, 2048]
  • blk.10.attn_v.weight
    Q8_0
    [1536, 2048]
  • blk.10.exp_probs_b.bias
    F32
    [128]
  • blk.10.ffn_down_exps.weight
    Q8_0
    [512, 1536, 128]
  • blk.10.ffn_down_shexp.weight
    Q8_0
    [512, 1536]
  • blk.10.ffn_gate_exps.weight
    Q8_0
    [1536, 512, 128]
  • blk.10.ffn_gate_inp.weight
    F32
    [1536, 128]
  • blk.10.ffn_gate_shexp.weight
    Q8_0
    [1536, 512]
  • blk.10.ffn_norm.weight
    F32
    [1536]
  • blk.10.ffn_up_exps.weight
    Q8_0
    [1536, 512, 128]
  • blk.10.ffn_up_shexp.weight
    Q8_0
    [1536, 512]
  • blk.10.ssm_a
    F32
    [1, 16]
  • blk.10.ssm_beta.weight
    Q8_0
    [1536, 16]
  • blk.10.ssm_conv1d_k.weight
    F32
    [4, 1, 2048]
  • blk.10.ssm_conv1d_q.weight
    F32
    [4, 1, 2048]
  • blk.10.ssm_conv1d_v.weight
    F32
    [4, 1, 2048]
  • blk.10.ssm_dt.bias
    F32
    [2048]
  • blk.10.ssm_f_a.weight
    Q8_0
    [1536, 2048]
  • blk.10.ssm_g_a.weight
    Q8_0
    [1536, 2048]
  • blk.10.ssm_norm.weight
    F32
    [128]
    • blk.11
  • blk.11.attn_gate.weight
    Q8_0
    [1536, 16]
  • blk.11.attn_k_b.weight
    Q8_0
    [128, 512, 16]
  • blk.11.attn_kv_a_mqa.weight
    Q8_0
    [1536, 576]
  • blk.11.attn_kv_a_norm.weight
    F32
    [512]
  • blk.11.attn_norm.weight
    F32
    [1536]
  • blk.11.attn_output.weight
    Q8_0
    [2048, 1536]
  • blk.11.attn_q_a.weight
    Q8_0
    [1536, 256]
  • blk.11.attn_q_a_norm.weight
    F32
    [256]
  • blk.11.attn_q_b.weight
    Q8_0
    [256, 3072]
  • blk.11.attn_v_b.weight
    Q8_0
    [512, 128, 16]
  • blk.11.exp_probs_b.bias
    F32
    [128]
  • blk.11.ffn_down_exps.weight
    Q8_0
    [512, 1536, 128]
  • blk.11.ffn_down_shexp.weight
    Q8_0
    [512, 1536]
  • blk.11.ffn_gate_exps.weight
    Q8_0
    [1536, 512, 128]
  • blk.11.ffn_gate_inp.weight
    F32
    [1536, 128]
  • blk.11.ffn_gate_shexp.weight
    Q8_0
    [1536, 512]
  • blk.11.ffn_norm.weight
    F32
    [1536]
  • blk.11.ffn_up_exps.weight
    Q8_0
    [1536, 512, 128]
  • blk.11.ffn_up_shexp.weight
    Q8_0
    [1536, 512]
    • blk.12
  • blk.12.attn_k.weight
    Q8_0
    [1536, 2048]
  • blk.12.attn_norm.weight
    F32
    [1536]
  • blk.12.attn_output.weight
    Q8_0
    [2048, 1536]
  • blk.12.attn_q.weight
    Q8_0
    [1536, 2048]
  • blk.12.attn_v.weight
    Q8_0
    [1536, 2048]
  • blk.12.exp_probs_b.bias
    F32
    [128]
  • blk.12.ffn_down_exps.weight
    Q8_0
    [512, 1536, 128]
  • blk.12.ffn_down_shexp.weight
    Q8_0
    [512, 1536]
  • blk.12.ffn_gate_exps.weight
    Q8_0
    [1536, 512, 128]
  • blk.12.ffn_gate_inp.weight
    F32
    [1536, 128]
  • blk.12.ffn_gate_shexp.weight
    Q8_0
    [1536, 512]
  • blk.12.ffn_norm.weight
    F32
    [1536]
  • blk.12.ffn_up_exps.weight
    Q8_0
    [1536, 512, 128]
  • blk.12.ffn_up_shexp.weight
    Q8_0
    [1536, 512]
  • blk.12.ssm_a
    F32
    [1, 16]
  • blk.12.ssm_beta.weight
    Q8_0
    [1536, 16]
  • blk.12.ssm_conv1d_k.weight
    F32
    [4, 1, 2048]
  • blk.12.ssm_conv1d_q.weight
    F32
    [4, 1, 2048]
  • blk.12.ssm_conv1d_v.weight
    F32
    [4, 1, 2048]
  • blk.12.ssm_dt.bias
    F32
    [2048]
  • blk.12.ssm_f_a.weight
    Q8_0
    [1536, 2048]
  • blk.12.ssm_g_a.weight
    Q8_0
    [1536, 2048]
  • blk.12.ssm_norm.weight
    F32
    [128]
    • blk.13
  • blk.13.attn_k.weight
    Q8_0
    [1536, 2048]
  • blk.13.attn_norm.weight
    F32
    [1536]
  • blk.13.attn_output.weight
    Q8_0
    [2048, 1536]
  • blk.13.attn_q.weight
    Q8_0
    [1536, 2048]
  • blk.13.attn_v.weight
    Q8_0
    [1536, 2048]
  • blk.13.exp_probs_b.bias
    F32
    [128]
  • blk.13.ffn_down_exps.weight
    Q8_0
    [512, 1536, 128]
  • blk.13.ffn_down_shexp.weight
    Q8_0
    [512, 1536]
  • blk.13.ffn_gate_exps.weight
    Q8_0
    [1536, 512, 128]
  • blk.13.ffn_gate_inp.weight
    F32
    [1536, 128]
  • blk.13.ffn_gate_shexp.weight
    Q8_0
    [1536, 512]
  • blk.13.ffn_norm.weight
    F32
    [1536]
  • blk.13.ffn_up_exps.weight
    Q8_0
    [1536, 512, 128]
  • blk.13.ffn_up_shexp.weight
    Q8_0
    [1536, 512]
  • blk.13.ssm_a
    F32
    [1, 16]
  • blk.13.ssm_beta.weight
    Q8_0
    [1536, 16]
  • blk.13.ssm_conv1d_k.weight
    F32
    [4, 1, 2048]
  • blk.13.ssm_conv1d_q.weight
    F32
    [4, 1, 2048]
  • blk.13.ssm_conv1d_v.weight
    F32
    [4, 1, 2048]
  • blk.13.ssm_dt.bias
    F32
    [2048]
  • blk.13.ssm_f_a.weight
    Q8_0
    [1536, 2048]
  • blk.13.ssm_g_a.weight
    Q8_0
    [1536, 2048]
  • blk.13.ssm_norm.weight
    F32
    [128]
    • blk.14
  • blk.14.attn_k.weight
    Q8_0
    [1536, 2048]
  • blk.14.attn_norm.weight
    F32
    [1536]
  • blk.14.attn_output.weight
    Q8_0
    [2048, 1536]
  • blk.14.attn_q.weight
    Q8_0
    [1536, 2048]
  • blk.14.attn_v.weight
    Q8_0
    [1536, 2048]
  • blk.14.exp_probs_b.bias
    F32
    [128]
  • blk.14.ffn_down_exps.weight
    Q8_0
    [512, 1536, 128]
  • blk.14.ffn_down_shexp.weight
    Q8_0
    [512, 1536]
  • blk.14.ffn_gate_exps.weight
    Q8_0
    [1536, 512, 128]
  • blk.14.ffn_gate_inp.weight
    F32
    [1536, 128]
  • blk.14.ffn_gate_shexp.weight
    Q8_0
    [1536, 512]
  • blk.14.ffn_norm.weight
    F32
    [1536]
  • blk.14.ffn_up_exps.weight
    Q8_0
    [1536, 512, 128]
  • blk.14.ffn_up_shexp.weight
    Q8_0
    [1536, 512]
  • blk.14.ssm_a
    F32
    [1, 16]
  • blk.14.ssm_beta.weight
    Q8_0
    [1536, 16]
  • blk.14.ssm_conv1d_k.weight
    F32
    [4, 1, 2048]
  • blk.14.ssm_conv1d_q.weight
    F32
    [4, 1, 2048]
  • blk.14.ssm_conv1d_v.weight
    F32
    [4, 1, 2048]
  • blk.14.ssm_dt.bias
    F32
    [2048]
  • blk.14.ssm_f_a.weight
    Q8_0
    [1536, 2048]
  • blk.14.ssm_g_a.weight
    Q8_0
    [1536, 2048]
  • blk.14.ssm_norm.weight
    F32
    [128]
    • blk.15
  • blk.15.attn_gate.weight
    Q8_0
    [1536, 16]
  • blk.15.attn_k_b.weight
    Q8_0
    [128, 512, 16]
  • blk.15.attn_kv_a_mqa.weight
    Q8_0
    [1536, 576]
  • blk.15.attn_kv_a_norm.weight
    F32
    [512]
  • blk.15.attn_norm.weight
    F32
    [1536]
  • blk.15.attn_output.weight
    Q8_0
    [2048, 1536]
  • blk.15.attn_q_a.weight
    Q8_0
    [1536, 256]
  • blk.15.attn_q_a_norm.weight
    F32
    [256]
  • blk.15.attn_q_b.weight
    Q8_0
    [256, 3072]
  • blk.15.attn_v_b.weight
    Q8_0
    [512, 128, 16]
  • blk.15.exp_probs_b.bias
    F32
    [128]
  • blk.15.ffn_down_exps.weight
    Q8_0
    [512, 1536, 128]
  • blk.15.ffn_down_shexp.weight
    Q8_0
    [512, 1536]
  • blk.15.ffn_gate_exps.weight
    Q8_0
    [1536, 512, 128]
  • blk.15.ffn_gate_inp.weight
    F32
    [1536, 128]
  • blk.15.ffn_gate_shexp.weight
    Q8_0
    [1536, 512]
  • blk.15.ffn_norm.weight
    F32
    [1536]
  • blk.15.ffn_up_exps.weight
    Q8_0
    [1536, 512, 128]
  • blk.15.ffn_up_shexp.weight
    Q8_0
    [1536, 512]
    • blk.16
  • blk.16.attn_k.weight
    Q8_0
    [1536, 2048]
  • blk.16.attn_norm.weight
    F32
    [1536]
  • blk.16.attn_output.weight
    Q8_0
    [2048, 1536]
  • blk.16.attn_q.weight
    Q8_0
    [1536, 2048]
  • blk.16.attn_v.weight
    Q8_0
    [1536, 2048]
  • blk.16.exp_probs_b.bias
    F32
    [128]
  • blk.16.ffn_down_exps.weight
    Q8_0
    [512, 1536, 128]
  • blk.16.ffn_down_shexp.weight
    Q8_0
    [512, 1536]
  • blk.16.ffn_gate_exps.weight
    Q8_0
    [1536, 512, 128]
  • blk.16.ffn_gate_inp.weight
    F32
    [1536, 128]
  • blk.16.ffn_gate_shexp.weight
    Q8_0
    [1536, 512]
  • blk.16.ffn_norm.weight
    F32
    [1536]
  • blk.16.ffn_up_exps.weight
    Q8_0
    [1536, 512, 128]
  • blk.16.ffn_up_shexp.weight
    Q8_0
    [1536, 512]
  • blk.16.ssm_a
    F32
    [1, 16]
  • blk.16.ssm_beta.weight
    Q8_0
    [1536, 16]
  • blk.16.ssm_conv1d_k.weight
    F32
    [4, 1, 2048]
  • blk.16.ssm_conv1d_q.weight
    F32
    [4, 1, 2048]
  • blk.16.ssm_conv1d_v.weight
    F32
    [4, 1, 2048]
  • blk.16.ssm_dt.bias
    F32
    [2048]
  • blk.16.ssm_f_a.weight
    Q8_0
    [1536, 2048]
  • blk.16.ssm_g_a.weight
    Q8_0
    [1536, 2048]
  • blk.16.ssm_norm.weight
    F32
    [128]
    • blk.17
  • blk.17.attn_k.weight
    Q8_0
    [1536, 2048]
  • blk.17.attn_norm.weight
    F32
    [1536]
  • blk.17.attn_output.weight
    Q8_0
    [2048, 1536]
  • blk.17.attn_q.weight
    Q8_0
    [1536, 2048]
  • blk.17.attn_v.weight
    Q8_0
    [1536, 2048]
  • blk.17.exp_probs_b.bias
    F32
    [128]
  • blk.17.ffn_down_exps.weight
    Q8_0
    [512, 1536, 128]
  • blk.17.ffn_down_shexp.weight
    Q8_0
    [512, 1536]
  • blk.17.ffn_gate_exps.weight
    Q8_0
    [1536, 512, 128]
  • blk.17.ffn_gate_inp.weight
    F32
    [1536, 128]
  • blk.17.ffn_gate_shexp.weight
    Q8_0
    [1536, 512]
  • blk.17.ffn_norm.weight
    F32
    [1536]
  • blk.17.ffn_up_exps.weight
    Q8_0
    [1536, 512, 128]
  • blk.17.ffn_up_shexp.weight
    Q8_0
    [1536, 512]
  • blk.17.ssm_a
    F32
    [1, 16]
  • blk.17.ssm_beta.weight
    Q8_0
    [1536, 16]
  • blk.17.ssm_conv1d_k.weight
    F32
    [4, 1, 2048]
  • blk.17.ssm_conv1d_q.weight
    F32
    [4, 1, 2048]
  • blk.17.ssm_conv1d_v.weight
    F32
    [4, 1, 2048]
  • blk.17.ssm_dt.bias
    F32
    [2048]
  • blk.17.ssm_f_a.weight
    Q8_0
    [1536, 2048]
  • blk.17.ssm_g_a.weight
    Q8_0
    [1536, 2048]
  • blk.17.ssm_norm.weight
    F32
    [128]
    • blk.18
  • blk.18.attn_k.weight
    Q8_0
    [1536, 2048]
  • blk.18.attn_norm.weight
    F32
    [1536]
  • blk.18.attn_output.weight
    Q8_0
    [2048, 1536]
  • blk.18.attn_q.weight
    Q8_0
    [1536, 2048]
  • blk.18.attn_v.weight
    Q8_0
    [1536, 2048]
  • blk.18.exp_probs_b.bias
    F32
    [128]
  • blk.18.ffn_down_exps.weight
    Q8_0
    [512, 1536, 128]
  • blk.18.ffn_down_shexp.weight
    Q8_0
    [512, 1536]
  • blk.18.ffn_gate_exps.weight
    Q8_0
    [1536, 512, 128]
  • blk.18.ffn_gate_inp.weight
    F32
    [1536, 128]
  • blk.18.ffn_gate_shexp.weight
    Q8_0
    [1536, 512]
  • blk.18.ffn_norm.weight
    F32
    [1536]
  • blk.18.ffn_up_exps.weight
    Q8_0
    [1536, 512, 128]
  • blk.18.ffn_up_shexp.weight
    Q8_0
    [1536, 512]
  • blk.18.ssm_a
    F32
    [1, 16]
  • blk.18.ssm_beta.weight
    Q8_0
    [1536, 16]
  • blk.18.ssm_conv1d_k.weight
    F32
    [4, 1, 2048]
  • blk.18.ssm_conv1d_q.weight
    F32
    [4, 1, 2048]
  • blk.18.ssm_conv1d_v.weight
    F32
    [4, 1, 2048]
  • blk.18.ssm_dt.bias
    F32
    [2048]
  • blk.18.ssm_f_a.weight
    Q8_0
    [1536, 2048]
  • blk.18.ssm_g_a.weight
    Q8_0
    [1536, 2048]
  • blk.18.ssm_norm.weight
    F32
    [128]
    • blk.19
  • blk.19.attn_gate.weight
    Q8_0
    [1536, 16]
  • blk.19.attn_k_b.weight
    Q8_0
    [128, 512, 16]
  • blk.19.attn_kv_a_mqa.weight
    Q8_0
    [1536, 576]
  • blk.19.attn_kv_a_norm.weight
    F32
    [512]
  • blk.19.attn_norm.weight
    F32
    [1536]
  • blk.19.attn_output.weight
    Q8_0
    [2048, 1536]
  • blk.19.attn_q_a.weight
    Q8_0
    [1536, 256]
  • blk.19.attn_q_a_norm.weight
    F32
    [256]
  • blk.19.attn_q_b.weight
    Q8_0
    [256, 3072]
  • blk.19.attn_v_b.weight
    Q8_0
    [512, 128, 16]
  • blk.19.exp_probs_b.bias
    F32
    [128]
  • blk.19.ffn_down_exps.weight
    Q8_0
    [512, 1536, 128]
  • blk.19.ffn_down_shexp.weight
    Q8_0
    [512, 1536]
  • blk.19.ffn_gate_exps.weight
    Q8_0
    [1536, 512, 128]
  • blk.19.ffn_gate_inp.weight
    F32
    [1536, 128]
  • blk.19.ffn_gate_shexp.weight
    Q8_0
    [1536, 512]
  • blk.19.ffn_norm.weight
    F32
    [1536]
  • blk.19.ffn_up_exps.weight
    Q8_0
    [1536, 512, 128]
  • blk.19.ffn_up_shexp.weight
    Q8_0
    [1536, 512]
    • blk.20
  • blk.20.attn_k.weight
    Q8_0
    [1536, 2048]
  • blk.20.attn_norm.weight
    F32
    [1536]
  • blk.20.attn_output.weight
    Q8_0
    [2048, 1536]
  • blk.20.attn_q.weight
    Q8_0
    [1536, 2048]
  • blk.20.attn_v.weight
    Q8_0
    [1536, 2048]
  • blk.20.exp_probs_b.bias
    F32
    [128]
  • blk.20.ffn_down_exps.weight
    Q8_0
    [512, 1536, 128]
  • blk.20.ffn_down_shexp.weight
    Q8_0
    [512, 1536]
  • blk.20.ffn_gate_exps.weight
    Q8_0
    [1536, 512, 128]
  • blk.20.ffn_gate_inp.weight
    F32
    [1536, 128]
  • blk.20.ffn_gate_shexp.weight
    Q8_0
    [1536, 512]
  • blk.20.ffn_norm.weight
    F32
    [1536]
  • blk.20.ffn_up_exps.weight
    Q8_0
    [1536, 512, 128]
  • blk.20.ffn_up_shexp.weight
    Q8_0
    [1536, 512]
  • blk.20.ssm_a
    F32
    [1, 16]
  • blk.20.ssm_beta.weight
    Q8_0
    [1536, 16]
  • blk.20.ssm_conv1d_k.weight
    F32
    [4, 1, 2048]
  • blk.20.ssm_conv1d_q.weight
    F32
    [4, 1, 2048]
  • blk.20.ssm_conv1d_v.weight
    F32
    [4, 1, 2048]
  • blk.20.ssm_dt.bias
    F32
    [2048]
  • blk.20.ssm_f_a.weight
    Q8_0
    [1536, 2048]
  • blk.20.ssm_g_a.weight
    Q8_0
    [1536, 2048]
  • blk.20.ssm_norm.weight
    F32
    [128]
    • blk.21
  • blk.21.attn_k.weight
    Q8_0
    [1536, 2048]
  • blk.21.attn_norm.weight
    F32
    [1536]
  • blk.21.attn_output.weight
    Q8_0
    [2048, 1536]
  • blk.21.attn_q.weight
    Q8_0
    [1536, 2048]
  • blk.21.attn_v.weight
    Q8_0
    [1536, 2048]
  • blk.21.exp_probs_b.bias
    F32
    [128]
  • blk.21.ffn_down_exps.weight
    Q8_0
    [512, 1536, 128]
  • blk.21.ffn_down_shexp.weight
    Q8_0
    [512, 1536]
  • blk.21.ffn_gate_exps.weight
    Q8_0
    [1536, 512, 128]
  • blk.21.ffn_gate_inp.weight
    F32
    [1536, 128]
  • blk.21.ffn_gate_shexp.weight
    Q8_0
    [1536, 512]
  • blk.21.ffn_norm.weight
    F32
    [1536]
  • blk.21.ffn_up_exps.weight
    Q8_0
    [1536, 512, 128]
  • blk.21.ffn_up_shexp.weight
    Q8_0
    [1536, 512]
  • blk.21.ssm_a
    F32
    [1, 16]
  • blk.21.ssm_beta.weight
    Q8_0
    [1536, 16]
  • blk.21.ssm_conv1d_k.weight
    F32
    [4, 1, 2048]
  • blk.21.ssm_conv1d_q.weight
    F32
    [4, 1, 2048]
  • blk.21.ssm_conv1d_v.weight
    F32
    [4, 1, 2048]
  • blk.21.ssm_dt.bias
    F32
    [2048]
  • blk.21.ssm_f_a.weight
    Q8_0
    [1536, 2048]
  • blk.21.ssm_g_a.weight
    Q8_0
    [1536, 2048]
  • blk.21.ssm_norm.weight
    F32
    [128]
    • blk.22
  • blk.22.attn_k.weight
    Q8_0
    [1536, 2048]
  • blk.22.attn_norm.weight
    F32
    [1536]
  • blk.22.attn_output.weight
    Q8_0
    [2048, 1536]
  • blk.22.attn_q.weight
    Q8_0
    [1536, 2048]
  • blk.22.attn_v.weight
    Q8_0
    [1536, 2048]
  • blk.22.exp_probs_b.bias
    F32
    [128]
  • blk.22.ffn_down_exps.weight
    Q8_0
    [512, 1536, 128]
  • blk.22.ffn_down_shexp.weight
    Q8_0
    [512, 1536]
  • blk.22.ffn_gate_exps.weight
    Q8_0
    [1536, 512, 128]
  • blk.22.ffn_gate_inp.weight
    F32
    [1536, 128]
  • blk.22.ffn_gate_shexp.weight
    Q8_0
    [1536, 512]
  • blk.22.ffn_norm.weight
    F32
    [1536]
  • blk.22.ffn_up_exps.weight
    Q8_0
    [1536, 512, 128]
  • blk.22.ffn_up_shexp.weight
    Q8_0
    [1536, 512]
  • blk.22.ssm_a
    F32
    [1, 16]
  • blk.22.ssm_beta.weight
    Q8_0
    [1536, 16]
  • blk.22.ssm_conv1d_k.weight
    F32
    [4, 1, 2048]
  • blk.22.ssm_conv1d_q.weight
    F32
    [4, 1, 2048]
  • blk.22.ssm_conv1d_v.weight
    F32
    [4, 1, 2048]
  • blk.22.ssm_dt.bias
    F32
    [2048]
  • blk.22.ssm_f_a.weight
    Q8_0
    [1536, 2048]
  • blk.22.ssm_g_a.weight
    Q8_0
    [1536, 2048]
  • blk.22.ssm_norm.weight
    F32
    [128]
    • blk.23
  • blk.23.attn_gate.weight
    Q8_0
    [1536, 16]
  • blk.23.attn_k_b.weight
    Q8_0
    [128, 512, 16]
  • blk.23.attn_kv_a_mqa.weight
    Q8_0
    [1536, 576]
  • blk.23.attn_kv_a_norm.weight
    F32
    [512]
  • blk.23.attn_norm.weight
    F32
    [1536]
  • blk.23.attn_output.weight
    Q8_0
    [2048, 1536]
  • blk.23.attn_q_a.weight
    Q8_0
    [1536, 256]
  • blk.23.attn_q_a_norm.weight
    F32
    [256]
  • blk.23.attn_q_b.weight
    Q8_0
    [256, 3072]
  • blk.23.attn_v_b.weight
    Q8_0
    [512, 128, 16]
  • blk.23.exp_probs_b.bias
    F32
    [128]
  • blk.23.ffn_down_exps.weight
    Q8_0
    [512, 1536, 128]
  • blk.23.ffn_down_shexp.weight
    Q8_0
    [512, 1536]
  • blk.23.ffn_gate_exps.weight
    Q8_0
    [1536, 512, 128]
  • blk.23.ffn_gate_inp.weight
    F32
    [1536, 128]
  • blk.23.ffn_gate_shexp.weight
    Q8_0
    [1536, 512]
  • blk.23.ffn_norm.weight
    F32
    [1536]
  • blk.23.ffn_up_exps.weight
    Q8_0
    [1536, 512, 128]
  • blk.23.ffn_up_shexp.weight
    Q8_0
    [1536, 512]
  • output.weight
    Q8_0
    [1536, 157184]
  • output_norm.weight
    F32
    [1536]