mlx-bun library API
    Preparing search index...

    Interface ModelRecord

    interface ModelRecord {
        expertsBytes: number;
        hasAudioConfig: boolean;
        hasAudioTower: boolean;
        hasKvConfig: boolean;
        hasToolTemplate: boolean;
        hasVisionSidecar: boolean;
        hiddenSize: number | null;
        license: string | null;
        modelType: string;
        numLayers: number | null;
        paramCount: number | null;
        path: string;
        quantBits: number | null;
        quantGroupSize: number | null;
        quantMode: string | null;
        repoId: string;
        scannedAt: number;
        sidecarBytes: number;
        sizeBytes: number;
        supportTier?: string | null;
        visionConfigType: string | null;
        vocabSize: number | null;
    }
    Index
    expertsBytes hasAudioConfig hasAudioTower hasKvConfig hasToolTemplate hasVisionSidecar hiddenSize license modelType numLayers paramCount path quantBits quantGroupSize quantMode repoId scannedAt sidecarBytes sizeBytes supportTier? visionConfigType vocabSize
    expertsBytes: number

    Bytes of .experts. tensors (MoE; 0 for dense models). Per-token decode only reads top_k/num_experts of these.

    hasAudioConfig: boolean

    config.json declares an audio_config block (gemma-4 Conformer).

    hasAudioTower: boolean

    The optiq_vision.safetensors sidecar actually ships audio_tower.* tensors (header-only check). Distinct from hasAudioConfig: the 12B OptiQ snapshot pairs an audio_config with a STUB sidecar holding only embed_audio.embedding_projection.weight, so config presence alone is not an audio-capability signal.

    hasKvConfig: boolean
    hasToolTemplate: boolean
    hasVisionSidecar: boolean
    hiddenSize: number | null
    license: string | null

    SPDX-ish license id from the model card's README frontmatter (e.g. "gemma" — custom terms — vs "apache-2.0"); null if absent. Surfaces license obligations per model in ls/fit output.

    modelType: string
    numLayers: number | null
    paramCount: number | null
    path: string
    quantBits: number | null
    quantGroupSize: number | null
    quantMode: string | null
    repoId: string
    scannedAt: number
    sidecarBytes: number

    optiq_vision.safetensors bytes (bf16 sidecar; loads only for vision).

    sizeBytes: number

    Language-model weight bytes (model-*.safetensors, sidecar excluded).

    supportTier?: string | null

    The support tier the caller's supportTier classifier assigned when the snapshot was indexed; absent when none did (or no classifier is set).

    visionConfigType: string | null

    config.json's vision_config.model_type when it names a vision tower (e.g. "gemma4_vision" — SigLIP, needs the bf16 sidecar — or "gemma4_unified_vision" — encoder-free). null for text-only models AND for configs whose nested vision_config is not a vision tower (Qwen3.5 nests a copy of its own text config there).

    vocabSize: number | null

    Settings
    Member Visibility

    On This Page
    Properties