41 1 month ago

The agent that finishes, faster. Onyx 2.1 says in one sentence what took five, starts answering in half the time, and still tells you what is still open. Reads code and images. 12B and 31B, 32K context.

vision tools thinking 12b 31b
675ad6e68101 · 175MB
    Metadata
  • general.architecture
    clip
  • general.file_type
    BF16
  • clip.audio.attention.head_count
    0
  • clip.audio.attention.layer_norm_epsilon
    1e-06
  • clip.audio.block_count
    0
  • clip.audio.embedding_length
    640
  • clip.audio.feed_forward_length
    0
  • clip.audio.num_mel_bins
    128
  • clip.audio.projection_dim
    3840
  • clip.audio.projector_type
    gemma4ua
  • clip.has_audio_encoder
    true
  • clip.has_vision_encoder
    true
  • clip.vision.attention.head_count
    0
  • clip.vision.attention.layer_norm_epsilon
    1e-06
  • clip.vision.block_count
    0
  • clip.vision.embedding_length
    3840
  • clip.vision.feed_forward_length
    0
  • clip.vision.image_mean
    [0, 0, 0]
  • clip.vision.image_size
    224
  • clip.vision.image_std
    [1, 1, 1]
  • clip.vision.patch_size
    16
  • clip.vision.projection_dim
    3840
  • clip.vision.projector_type
    gemma4uv
  • Tensor
  • mm.a.input_projection.weight
    BF16
    [640, 3840]
  • mm.input_projection.weight
    BF16
    [3840, 3840]
  • v.patch_embd.bias
    F32
    [3840]
  • v.patch_embd.weight
    F32
    [6912, 3840]
  • v.patch_norm.1.bias
    F32
    [6912]
  • v.patch_norm.1.weight
    F32
    [6912]
  • v.patch_norm.2.bias
    F32
    [3840]
  • v.patch_norm.2.weight
    F32
    [3840]
  • v.patch_norm.3.bias
    F32
    [3840]
  • v.patch_norm.3.weight
    F32
    [3840]
  • v.position_embd.weight
    F32
    [3840, 1120, 2]