mannix/ gemma4-98e-v6-coder:mtp-Q4_K_M

1,451 3 months ago

The best 20b coding model just got better! Beats the bigger 26b brother in Python and code reasoning

vision tools thinking
... /
draft
2a7bc8ed3f50 ยท 446MB
    Metadata
  • general.architecture
    gemma4_assistant
  • general.file_type
    Q8_0
  • gemma4_assistant.attention.head_count
    16
  • gemma4_assistant.attention.head_count_kv
    [8, 8, 8, 2]
  • gemma4_assistant.attention.k_eq_v
    true
  • gemma4_assistant.attention.key_length
    512
  • gemma4_assistant.attention.key_length_swa
    256
  • gemma4_assistant.attention.layer_norm_rms_epsilon
    1e-06
  • gemma4_assistant.attention.shared_kv_layers
    4
  • gemma4_assistant.attention.sliding_window
    1024
  • gemma4_assistant.attention.sliding_window_pattern
    [true, true, true, false]
  • gemma4_assistant.attention.value_length
    512
  • gemma4_assistant.attention.value_length_swa
    256
  • gemma4_assistant.block_count
    4
  • gemma4_assistant.centroid_top_k
    32
  • gemma4_assistant.context_length
    262144
  • gemma4_assistant.embedding_length
    1024
  • gemma4_assistant.embedding_length_per_layer_input
    0
  • gemma4_assistant.feed_forward_length
    8192
  • gemma4_assistant.n_centroids
    2048
  • gemma4_assistant.n_embd_backbone
    2816
  • gemma4_assistant.requires_target_arch
    gemma4
  • gemma4_assistant.rope.dimension_count
    512
  • gemma4_assistant.rope.dimension_count_swa
    256
  • gemma4_assistant.rope.freq_base
    1e+06
  • gemma4_assistant.rope.freq_base_swa
    10000
  • gemma4_assistant.use_ordered_embeddings
    false
  • gemma4_assistant.vocab_size
    262144
  • tokenizer.ggml.model
    none
    • Tensor
  • token_embd.weight
    Q8_0
    [1024, 262144]
    • blk.0
  • blk.0.attn_norm.weight
    F32
    [1024]
  • blk.0.attn_output.weight
    Q8_0
    [4096, 1024]
  • blk.0.attn_q.weight
    Q8_0
    [1024, 4096]
  • blk.0.attn_q_norm.weight
    F32
    [256]
  • blk.0.ffn_down.weight
    Q8_0
    [8192, 1024]
  • blk.0.ffn_gate.weight
    Q8_0
    [1024, 8192]
  • blk.0.ffn_norm.weight
    F32
    [1024]
  • blk.0.ffn_up.weight
    Q8_0
    [1024, 8192]
  • blk.0.layer_output_scale.weight
    F32
    [1]
  • blk.0.post_attention_norm.weight
    F32
    [1024]
  • blk.0.post_ffw_norm.weight
    F32
    [1024]
    • blk.1
  • blk.1.attn_norm.weight
    F32
    [1024]
  • blk.1.attn_output.weight
    Q8_0
    [4096, 1024]
  • blk.1.attn_q.weight
    Q8_0
    [1024, 4096]
  • blk.1.attn_q_norm.weight
    F32
    [256]
  • blk.1.ffn_down.weight
    Q8_0
    [8192, 1024]
  • blk.1.ffn_gate.weight
    Q8_0
    [1024, 8192]
  • blk.1.ffn_norm.weight
    F32
    [1024]
  • blk.1.ffn_up.weight
    Q8_0
    [1024, 8192]
  • blk.1.layer_output_scale.weight
    F32
    [1]
  • blk.1.post_attention_norm.weight
    F32
    [1024]
  • blk.1.post_ffw_norm.weight
    F32
    [1024]
    • blk.2
  • blk.2.attn_norm.weight
    F32
    [1024]
  • blk.2.attn_output.weight
    Q8_0
    [4096, 1024]
  • blk.2.attn_q.weight
    Q8_0
    [1024, 4096]
  • blk.2.attn_q_norm.weight
    F32
    [256]
  • blk.2.ffn_down.weight
    Q8_0
    [8192, 1024]
  • blk.2.ffn_gate.weight
    Q8_0
    [1024, 8192]
  • blk.2.ffn_norm.weight
    F32
    [1024]
  • blk.2.ffn_up.weight
    Q8_0
    [1024, 8192]
  • blk.2.layer_output_scale.weight
    F32
    [1]
  • blk.2.post_attention_norm.weight
    F32
    [1024]
  • blk.2.post_ffw_norm.weight
    F32
    [1024]
    • blk.3
  • blk.3.attn_norm.weight
    F32
    [1024]
  • blk.3.attn_output.weight
    Q8_0
    [8192, 1024]
  • blk.3.attn_q.weight
    Q8_0
    [1024, 8192]
  • blk.3.attn_q_norm.weight
    F32
    [512]
  • blk.3.ffn_down.weight
    Q8_0
    [8192, 1024]
  • blk.3.ffn_gate.weight
    Q8_0
    [1024, 8192]
  • blk.3.ffn_norm.weight
    F32
    [1024]
  • blk.3.ffn_up.weight
    Q8_0
    [1024, 8192]
  • blk.3.layer_output_scale.weight
    F32
    [1]
  • blk.3.post_attention_norm.weight
    F32
    [1024]
  • blk.3.post_ffw_norm.weight
    F32
    [1024]
  • mtp.post_projection.weight
    Q8_0
    [1024, 2816]
  • mtp.pre_projection.weight
    Q8_0
    [5632, 1024]
  • rope_freqs.weight
    F32
    [256]
  • output_norm.weight
    F32
    [1024]