CherryHQ/cherry-studio · critical · Error

Local model inference is not supported on Intel Mac (darwin

Error message

Local model inference is not supported on Intel Mac (darwin x64) — onnxruntime-node ships no darwin-x64 binding.

What it means

Thrown by InferenceServiceBase.ensureWorker() as a hard platform guard on Intel Mac (darwin x64). The onnxruntime-node package ships no native binding for darwin-x64, so spawning a worker would crash immediately on import. This check is the 'last line of defense' — the settings UI and knowledge-base cards already hide local-model features on Intel Mac, but this catches any programmatic call (embed, loadEmbedding, recognize, OCR agent tool) that bypassed the UI gate.

Source

Thrown at src/main/ai/inference/InferenceServiceBase.ts:108

  private closing = false
  private readonly logger: ReturnType<typeof loggerService.withContext>

  protected constructor(kind: LocalModelKind) {
    super()
    this.logger = loggerService.withContext(`InferenceService:${kind}`)
  }

  private async ensureWorker(): Promise<Worker> {
    if (this.closing) {
      throw new Error('inference host is shutting down')
    }
    // Last line of defense: the settings/KB cards already hide on Intel Mac (see
    // LocalModelDownloadService.getStatus), but this is the spawn point every
    // caller (embed/loadEmbedding/recognize, including the OCR agent tool)
    // funnels through, so anything that reaches it programmatically fails fast
    // instead of loading a worker that will crash on the missing native binding.
    if (isDarwinX64) {
      throw new Error(
        'Local model inference is not supported on Intel Mac (darwin x64) — onnxruntime-node ships no darwin-x64 binding.'
      )
    }
    const generation = this.workerGeneration
    const proxyRouting = await application.get('ProxyService').getRoutingSnapshot()
    const runtimeProfile = resolveLocalInferenceProfile(
      application.get('PreferenceService').get('feature.local_model.hardware_acceleration.enabled')
    )
    if (generation !== this.workerGeneration) {
      throw new Error('inference host terminated')
    }
    if (this.closing) {
      throw new Error('inference host is shutting down')
    }
    if (this.worker && this.workerProxyVersion === proxyRouting.version && this.workerProfileId === runtimeProfile.id) {
      return this.worker
    }
    if (this.worker) {

View on GitHub (pinned to 726446b54c)

Solutions

1. Switch to an Apple Silicon Mac (arm64) where onnxruntime-node has a working native binding.
2. Use a cloud-based embedding/OCR provider instead of local model inference on Intel Mac.
3. If running under Rosetta 2, run the native arm64 binary of Electron/Node instead of the x64 version.
4. Ensure the UI gating (LocalModelDownloadService.getStatus) is not bypassed by programmatic callers on Intel Mac.

import { isDarwinX64 } from '@main/core/platform'

// Gate all inference calls on platform support
if (isDarwinX64) {
  logger.info('Local model inference is not available on Intel Mac — use a cloud provider')
  return null // or fall back to a cloud-based embedding/OCR service
}

// Check platform support before any inference call
function supportsLocalInference(): boolean {
  return !isDarwinX64 // true on arm64 Mac, Linux, Windows; false on Intel Mac
}

When it happens

Trigger: Called on every inference request path: EmbeddingInferenceService.embed()/loadEmbedding() and OcrInferenceService.recognize() all funnel through ensureWorker(). On darwin x64, the isDarwinX64 constant (from @main/core/platform) is true, so the error fires immediately before any worker spawn attempt.

Common situations: Running Cherry Studio on an Intel Mac (pre-Apple-Silicon) and a knowledge-base embedding or OCR task was triggered programmatically without going through the UI that hides these features; a corrupted or misdetected platform check; running under Rosetta 2 which may report as x64.

AI-assisted analysis of CherryHQ/cherry-studio@726446b54c (2026-08-12). Data as JSON: /api/errors/4e869a320a7afa8d. Report an issue: GitHub.