recipe:
  name: solar-open2-250b-mixed-v1
  source_model: upstage/Solar-Open2-250B
  source_revision: 9190fbe63a2ad8e17fc766ccceb36de7c66f004b
  projected_size_gib: 88.97161643393338
  hard_limit_gib: 92
  requires_imatrix: true
  publication_candidate: true
  precision_schedule_locked: true

targets:
  token_embedding: Q8_0
  lm_head: Q8_0
  norms: F32
  router_control: F32
  gqa_projections: Q8_0
  kda_matrix_projections: Q8_0
  kda_small_control_norm_conv: F32
  shared_expert: Q8_0
  routed_expert_gate_up: IQ2_XXS
  routed_expert_down: Q3_K
  edge_routed_layers_0_3_44_47: Q4_K

protected_until_last:
  - kda_matrix_projections
  - kda_small_control_norm_conv
  - gqa_projections
  - router_control
  - shared_expert
  - token_embedding
  - lm_head

notes:
  - The 88.9716 GiB exact-layout projection is accepted. Do not lower any
    target precision or generate alternate lower-capacity mixed candidates.
  - Small KDA controls, gated norm, and depthwise convolution tensors are kept
    in F32 because GGML block quantization is either inapplicable to their row
    width or saves negligible capacity. This is a higher-precision fallback.