{
  "schema_version": "0.1.0",
  "generated_at": "2026-07-15",
  "target_ram_gb": 16,
  "variants": [
    {
      "variant_id": "prism-ml/Bonsai-1.7B-unpacked",
      "family": "Bonsai",
      "source_model_id": "Qwen/Qwen3-1.7B",
      "category": "llm",
      "total_parameters_b": 1.7,
      "format": "safetensors-fp16",
      "bits_per_weight": 16,
      "memory_estimate_gb": 3.4,
      "memory_estimate_basis": "derived-weight-only",
      "fit_status": "yes",
      "license": "apache-2.0",
      "license_status": "needs-card-review",
      "official_checkpoint": true,
      "runtime_support": ["transformers"],
      "source_url": "https://huggingface.co/prism-ml/Bonsai-1.7B-unpacked",
      "last_verified": "2026-07-15",
      "notes": "Official Prism ML unpacked checkpoint; retain Qwen3 base parameter identity and do not count this as a new base family."
    },
    {
      "variant_id": "prism-ml/Bonsai-4B-unpacked",
      "family": "Bonsai",
      "source_model_id": "Qwen/Qwen3-4B",
      "category": "llm",
      "total_parameters_b": 4.0,
      "format": "safetensors-fp16",
      "bits_per_weight": 16,
      "memory_estimate_gb": 8.0,
      "memory_estimate_basis": "derived-weight-only",
      "fit_status": "yes",
      "license": "apache-2.0",
      "license_status": "needs-card-review",
      "official_checkpoint": true,
      "runtime_support": ["transformers"],
      "source_url": "https://huggingface.co/prism-ml/Bonsai-4B-unpacked",
      "last_verified": "2026-07-15",
      "notes": "Official Prism ML unpacked checkpoint; retain Qwen3 base parameter identity and do not count this as a new base family."
    },
    {
      "variant_id": "prism-ml/Bonsai-8B-unpacked",
      "family": "Bonsai",
      "source_model_id": "Qwen/Qwen3-8B",
      "category": "llm",
      "total_parameters_b": 8.0,
      "format": "safetensors-fp16",
      "bits_per_weight": 16,
      "memory_estimate_gb": 16.0,
      "memory_estimate_basis": "derived-weight-only",
      "fit_status": "borderline",
      "license": "apache-2.0",
      "license_status": "needs-card-review",
      "official_checkpoint": true,
      "runtime_support": ["transformers"],
      "source_url": "https://huggingface.co/prism-ml/Bonsai-8B-unpacked",
      "last_verified": "2026-07-15",
      "notes": "Weight-only estimate consumes the full 16GB budget before runtime overhead; not recommended for a 16GB device without quantization."
    },
    {
      "variant_id": "prism-ml/Ternary-Bonsai-8B-unpacked",
      "family": "Ternary Bonsai",
      "source_model_id": "Qwen/Qwen3-8B",
      "category": "llm",
      "total_parameters_b": 8.0,
      "format": "safetensors-fp16-unpacked",
      "bits_per_weight": 16,
      "memory_estimate_gb": 16.0,
      "memory_estimate_basis": "official-card-weight-only",
      "fit_status": "borderline",
      "license": "apache-2.0",
      "license_status": "verified",
      "official_checkpoint": true,
      "runtime_support": ["transformers"],
      "source_url": "https://huggingface.co/prism-ml/Ternary-Bonsai-8B-unpacked",
      "last_verified": "2026-07-15",
      "notes": "The official card describes this as an unpacked FP16 file and explicitly says the packed format is where the memory benefit comes from."
    },
    {
      "variant_id": "prism-ml/Bonsai-1.7B-mlx-1bit",
      "family": "Bonsai",
      "source_model_id": "Qwen/Qwen3-1.7B",
      "category": "llm",
      "total_parameters_b": 1.7,
      "format": "mlx-1bit",
      "bits_per_weight": 1.25,
      "memory_estimate_gb": 0.27,
      "memory_estimate_basis": "derived-packed-weight-estimate",
      "fit_status": "yes",
      "license": "apache-2.0",
      "license_status": "needs-card-review",
      "official_checkpoint": true,
      "runtime_support": ["mlx"],
      "source_url": "https://huggingface.co/prism-ml/Bonsai-1.7B-mlx-1bit",
      "last_verified": "2026-07-15",
      "notes": "Packed Apple Silicon deployment variant; total parameters remain 1.7B despite the compressed footprint."
    },
    {
      "variant_id": "prism-ml/Bonsai-4B-mlx-1bit",
      "family": "Bonsai",
      "source_model_id": "Qwen/Qwen3-4B",
      "category": "llm",
      "total_parameters_b": 4.0,
      "format": "mlx-1bit",
      "bits_per_weight": 1.25,
      "memory_estimate_gb": 0.63,
      "memory_estimate_basis": "official-card-weight-only",
      "fit_status": "yes",
      "license": "apache-2.0",
      "license_status": "verified",
      "official_checkpoint": true,
      "runtime_support": ["mlx"],
      "source_url": "https://huggingface.co/prism-ml/Bonsai-4B-mlx-1bit",
      "last_verified": "2026-07-15",
      "notes": "Official card reports a 0.63GB deployed weight footprint and Apple Silicon/iPhone support."
    },
    {
      "variant_id": "prism-ml/Bonsai-8B-mlx-1bit",
      "family": "Bonsai",
      "source_model_id": "Qwen/Qwen3-8B",
      "category": "llm",
      "total_parameters_b": 8.0,
      "format": "mlx-1bit",
      "bits_per_weight": 1.25,
      "memory_estimate_gb": 1.26,
      "memory_estimate_basis": "derived-packed-weight-estimate",
      "fit_status": "yes",
      "license": "apache-2.0",
      "license_status": "needs-card-review",
      "official_checkpoint": true,
      "runtime_support": ["mlx"],
      "source_url": "https://huggingface.co/prism-ml/Bonsai-8B-mlx-1bit",
      "last_verified": "2026-07-15",
      "notes": "Packed Apple Silicon deployment variant; total parameters remain 8B despite the compressed footprint."
    },
    {
      "variant_id": "prism-ml/Ternary-Bonsai-1.7B-mlx-2bit",
      "family": "Ternary Bonsai",
      "source_model_id": "Qwen/Qwen3-1.7B",
      "category": "llm",
      "total_parameters_b": 1.7,
      "format": "mlx-2bit",
      "bits_per_weight": 2.0,
      "memory_estimate_gb": 0.43,
      "memory_estimate_basis": "derived-packed-weight-estimate",
      "fit_status": "yes",
      "license": "apache-2.0",
      "license_status": "needs-card-review",
      "official_checkpoint": true,
      "runtime_support": ["mlx"],
      "source_url": "https://huggingface.co/prism-ml/Ternary-Bonsai-1.7B-mlx-2bit",
      "last_verified": "2026-07-15",
      "notes": "Packed ternary deployment variant; not a separate base model."
    },
    {
      "variant_id": "prism-ml/Ternary-Bonsai-4B-mlx-2bit",
      "family": "Ternary Bonsai",
      "source_model_id": "Qwen/Qwen3-4B",
      "category": "llm",
      "total_parameters_b": 4.0,
      "format": "mlx-2bit",
      "bits_per_weight": 2.0,
      "memory_estimate_gb": 1.0,
      "memory_estimate_basis": "derived-packed-weight-estimate",
      "fit_status": "yes",
      "license": "apache-2.0",
      "license_status": "needs-card-review",
      "official_checkpoint": true,
      "runtime_support": ["mlx"],
      "source_url": "https://huggingface.co/prism-ml/Ternary-Bonsai-4B-mlx-2bit",
      "last_verified": "2026-07-15",
      "notes": "Packed ternary deployment variant; not a separate base model."
    },
    {
      "variant_id": "prism-ml/Ternary-Bonsai-8B-mlx-2bit",
      "family": "Ternary Bonsai",
      "source_model_id": "Qwen/Qwen3-8B",
      "category": "llm",
      "total_parameters_b": 8.0,
      "format": "mlx-2bit",
      "bits_per_weight": 2.0,
      "memory_estimate_gb": 2.0,
      "memory_estimate_basis": "derived-packed-weight-estimate",
      "fit_status": "yes",
      "license": "apache-2.0",
      "license_status": "needs-card-review",
      "official_checkpoint": true,
      "runtime_support": ["mlx"],
      "source_url": "https://huggingface.co/prism-ml/Ternary-Bonsai-8B-mlx-2bit",
      "last_verified": "2026-07-15",
      "notes": "Packed ternary deployment variant; not a separate base model."
    },
    {
      "variant_id": "prism-ml/Bonsai-1.7B-gguf",
      "family": "Bonsai",
      "source_model_id": "Qwen/Qwen3-1.7B",
      "category": "llm",
      "total_parameters_b": 1.7,
      "format": "gguf",
      "bits_per_weight": null,
      "memory_estimate_gb": null,
      "memory_estimate_basis": "requires-quantization-level",
      "fit_status": "unknown",
      "license": "apache-2.0",
      "license_status": "needs-card-review",
      "official_checkpoint": true,
      "runtime_support": ["llama_cpp"],
      "source_url": "https://huggingface.co/prism-ml/Bonsai-1.7B-gguf",
      "last_verified": "2026-07-15",
      "notes": "GGUF family record; exact memory depends on the selected quantization file."
    },
    {
      "variant_id": "prism-ml/Bonsai-4B-gguf",
      "family": "Bonsai",
      "source_model_id": "Qwen/Qwen3-4B",
      "category": "llm",
      "total_parameters_b": 4.0,
      "format": "gguf",
      "bits_per_weight": null,
      "memory_estimate_gb": null,
      "memory_estimate_basis": "requires-quantization-level",
      "fit_status": "unknown",
      "license": "apache-2.0",
      "license_status": "needs-card-review",
      "official_checkpoint": true,
      "runtime_support": ["llama_cpp"],
      "source_url": "https://huggingface.co/prism-ml/Bonsai-4B-gguf",
      "last_verified": "2026-07-15",
      "notes": "GGUF family record; exact memory depends on the selected quantization file."
    },
    {
      "variant_id": "prism-ml/Bonsai-8B-gguf",
      "family": "Bonsai",
      "source_model_id": "Qwen/Qwen3-8B",
      "category": "llm",
      "total_parameters_b": 8.0,
      "format": "gguf",
      "bits_per_weight": null,
      "memory_estimate_gb": null,
      "memory_estimate_basis": "requires-quantization-level",
      "fit_status": "unknown",
      "license": "apache-2.0",
      "license_status": "needs-card-review",
      "official_checkpoint": true,
      "runtime_support": ["llama_cpp"],
      "source_url": "https://huggingface.co/prism-ml/Bonsai-8B-gguf",
      "last_verified": "2026-07-15",
      "notes": "GGUF family record; exact memory depends on the selected quantization file."
    },
    {
      "variant_id": "prism-ml/bonsai-image-binary-4B-unpacked",
      "family": "Bonsai Image",
      "source_model_id": null,
      "category": "image-generation",
      "total_parameters_b": 4.0,
      "format": "safetensors-fp16",
      "bits_per_weight": 16,
      "memory_estimate_gb": 8.0,
      "memory_estimate_basis": "derived-weight-only",
      "fit_status": "yes",
      "license": "apache-2.0",
      "license_status": "needs-card-review",
      "official_checkpoint": true,
      "runtime_support": ["mlx"],
      "source_url": "https://huggingface.co/prism-ml/bonsai-image-binary-4B-unpacked",
      "last_verified": "2026-07-15",
      "notes": "Image-generation deployment variant; kept outside the language-model count."
    },
    {
      "variant_id": "prism-ml/bonsai-image-ternary-4B-unpacked",
      "family": "Bonsai Image",
      "source_model_id": null,
      "category": "image-generation",
      "total_parameters_b": 4.0,
      "format": "safetensors-fp16-unpacked",
      "bits_per_weight": 16,
      "memory_estimate_gb": 8.0,
      "memory_estimate_basis": "derived-weight-only",
      "fit_status": "yes",
      "license": "apache-2.0",
      "license_status": "needs-card-review",
      "official_checkpoint": true,
      "runtime_support": ["mlx"],
      "source_url": "https://huggingface.co/prism-ml/bonsai-image-ternary-4B-unpacked",
      "last_verified": "2026-07-15",
      "notes": "Image-generation deployment variant; kept outside the language-model count."
    },
    {
      "variant_id": "prism-ml/bonsai-image-binary-4B-mlx-1bit",
      "family": "Bonsai Image",
      "source_model_id": null,
      "category": "image-generation",
      "total_parameters_b": 4.0,
      "format": "mlx-1bit",
      "bits_per_weight": 1.25,
      "memory_estimate_gb": 0.63,
      "memory_estimate_basis": "derived-packed-weight-estimate",
      "fit_status": "yes",
      "license": "apache-2.0",
      "license_status": "needs-card-review",
      "official_checkpoint": true,
      "runtime_support": ["mlx"],
      "source_url": "https://huggingface.co/prism-ml/bonsai-image-binary-4B-mlx-1bit",
      "last_verified": "2026-07-15",
      "notes": "Packed image-generation variant; kept outside the language-model count."
    },
    {
      "variant_id": "prism-ml/bonsai-image-ternary-4B-mlx-2bit",
      "family": "Bonsai Image",
      "source_model_id": null,
      "category": "image-generation",
      "total_parameters_b": 4.0,
      "format": "mlx-2bit",
      "bits_per_weight": 2.0,
      "memory_estimate_gb": 1.0,
      "memory_estimate_basis": "derived-packed-weight-estimate",
      "fit_status": "yes",
      "license": "apache-2.0",
      "license_status": "needs-card-review",
      "official_checkpoint": true,
      "runtime_support": ["mlx"],
      "source_url": "https://huggingface.co/prism-ml/bonsai-image-ternary-4B-mlx-2bit",
      "last_verified": "2026-07-15",
      "notes": "Packed image-generation variant; kept outside the language-model count."
    },
    {
      "variant_id": "prism-ml/Bonsai-27B-mlx-1bit",
      "family": "Bonsai",
      "source_model_id": "Qwen/Qwen3.6-27B",
      "category": "llm",
      "parameter_scope": "edge-exception",
      "total_parameters_b": 27.76,
      "format": "mlx-1bit",
      "bits_per_weight": 1.25,
      "memory_estimate_gb": 12.2,
      "memory_estimate_basis": "official-card-peak-at-100k-context",
      "fit_status": "yes",
      "license": "apache-2.0",
      "license_status": "verified",
      "official_checkpoint": true,
      "runtime_support": ["mlx", "mlx-swift"],
      "source_url": "https://huggingface.co/prism-ml/Bonsai-27B-mlx-1bit",
      "last_verified": "2026-07-15",
      "notes": "27.3B binary language weights plus a 0.46B vision tower. Official MLX peak is 12.2GB at 100K context; this is included as an edge-deployment exception and not in the primary sub-12B count."
    },
    {
      "variant_id": "prism-ml/Bonsai-27B-gguf",
      "family": "Bonsai",
      "source_model_id": "Qwen/Qwen3.6-27B",
      "category": "llm",
      "parameter_scope": "edge-exception",
      "total_parameters_b": 27.76,
      "format": "gguf-q1_0_g128",
      "bits_per_weight": 1.125,
      "memory_estimate_gb": 11.6,
      "memory_estimate_basis": "official-card-peak-at-100k-context",
      "fit_status": "yes",
      "license": "apache-2.0",
      "license_status": "verified",
      "official_checkpoint": true,
      "runtime_support": ["llama_cpp"],
      "source_url": "https://huggingface.co/prism-ml/Bonsai-27B-gguf",
      "last_verified": "2026-07-15",
      "notes": "Official 1-bit GGUF artifact with custom llama.cpp kernels. The card reports 11.6GB peak at 100K context; this is an edge-deployment exception and not in the primary sub-12B count."
    },
    {
      "variant_id": "prism-ml/Ternary-Bonsai-27B-mlx-2bit",
      "family": "Ternary Bonsai",
      "source_model_id": "Qwen/Qwen3.6-27B",
      "category": "llm",
      "parameter_scope": "edge-exception",
      "total_parameters_b": 27.76,
      "format": "mlx-2bit",
      "bits_per_weight": 2.25,
      "memory_estimate_gb": 15.5,
      "memory_estimate_basis": "official-card-peak-at-100k-context",
      "fit_status": "yes",
      "license": "apache-2.0",
      "license_status": "verified",
      "official_checkpoint": true,
      "runtime_support": ["mlx"],
      "source_url": "https://huggingface.co/prism-ml/Ternary-Bonsai-27B-mlx-2bit",
      "last_verified": "2026-07-15",
      "notes": "27.3B ternary language weights plus a 0.46B vision tower. The MLX pack uses 2.25 effective bits/weight and the official card reports 15.5GB peak at 100K context; this is an edge-deployment exception and not in the primary sub-12B count."
    },
    {
      "variant_id": "prism-ml/Ternary-Bonsai-27B-gguf",
      "family": "Ternary Bonsai",
      "source_model_id": "Qwen/Qwen3.6-27B",
      "category": "llm",
      "parameter_scope": "edge-exception",
      "total_parameters_b": 27.76,
      "format": "gguf-q2_0_g128",
      "bits_per_weight": 1.71,
      "memory_estimate_gb": 14.7,
      "memory_estimate_basis": "official-card-peak-at-100k-context",
      "fit_status": "yes",
      "license": "apache-2.0",
      "license_status": "verified",
      "official_checkpoint": true,
      "runtime_support": ["llama_cpp"],
      "source_url": "https://huggingface.co/prism-ml/Ternary-Bonsai-27B-gguf",
      "last_verified": "2026-07-15",
      "notes": "Official ternary GGUF artifact with custom llama.cpp kernels. The card reports 14.7GB peak at 100K context; this is an edge-deployment exception and not in the primary sub-12B count."
    }
  ]
}