Skip to content

v1.3.2 — Apple Intelligence removed (empirically unusable)

Choose a tag to compare

@mosandlt mosandlt released this 18 Apr 17:59
· 20 commits to main since this release

🇬🇧 English below · 🇩🇪 Deutsch darunter


🇬🇧 English

Apple Intelligence integration (v1.3.0 / v1.3.1) removed.

After live-testing all six LLM-backed modes with real prompts, the on-device ~3B foundation model failed every one of them:

  • Ignores the system prompt (Rage → generic assistant text, not a rewrite)
  • Hallucinates fluently on unusual inputs (Samsung-TV-via-toaster → fabricated Python script with invalid API calls)
  • Loops on repetitive output
  • Rewrites content in Emoji mode instead of preserving it 1:1

A "local" provider whose only functional role is to route every call to a cloud fallback is theater. Removed entirely rather than ship the mirage.

Kept from v1.3.1

Privacy Mode auto-skips for Ollama. That's a genuinely useful, provider-independent improvement — when the active profile is Ollama, the pre-send anonymizer short-circuits (data stays on your Mac, no over-anonymization of code identifiers like meineFunktion / file.swift). UI toggle unchanged; log line Privacy: skipped (local provider: ollama) when it fires.

If you want a larger local model

Use Ollama — it's been supported since v1.0.7:

ollama pull qwen2.5:14b      # ~10 GB RAM
ollama pull qwen2.5:32b      # ~22 GB RAM
ollama pull llama3.3:70b     # ~45 GB, M4 Max/Ultra territory

Then in Settings → Profile → New profile: Ollama / http://localhost:11434 / pick the model you pulled.

Migration

If you had a v1.3.0 Apple-Intelligence profile saved, it persists in UserDefaults but no longer appears in the provider picker. Delete it manually in Settings → Profile. No other migration needed.

Install

  1. Download blitzbot-1.3.2-macos-arm64.zip
  2. Unzip, drag blitzbot.app to /Applications
  3. First launch: right-click → Open
  4. macOS 15+ flagging damaged? xattr -dr com.apple.quarantine /Applications/blitzbot.app

🇩🇪 Deutsch

Apple-Intelligence-Integration (v1.3.0 / v1.3.1) entfernt.

Nach Live-Tests aller sechs LLM-Modi mit realen Prompts versagt das on-device ~3B Foundation Model in jedem einzelnen:

  • Ignoriert den System-Prompt (Rage → generischer Assistenten-Text statt Umschrift)
  • Halluziniert flüssig bei ungewöhnlichen Inputs (Samsung-TV-via-Toaster → erfundenes Python-Skript mit ungültigen API-Calls)
  • Loopt bei wiederholenden Outputs
  • Schreibt Content in Emoji-Modus um statt ihn 1:1 zu erhalten

Ein „lokaler" Provider, dessen einzige funktionale Rolle ist, jeden Call an einen Cloud-Fallback weiterzuleiten, ist Theater. Komplett entfernt statt die Attrappe auszuliefern.

Aus v1.3.1 behalten

Privacy-Mode überspringt Ollama automatisch. Das ist eine echte, provider-unabhängige Verbesserung — wenn das aktive Profil Ollama ist, schaltet der Pre-Send-Anonymizer ab (Daten bleiben auf deinem Mac, keine Over-Anonymisierung von Code-Identifiern wie meineFunktion / file.swift). UI-Toggle unverändert; Log-Zeile Privacy: skipped (local provider: ollama) wenn es greift.

Falls du ein größeres lokales Modell willst

Nutze Ollama — ist seit v1.0.7 unterstützt:

ollama pull qwen2.5:14b      # ~10 GB RAM
ollama pull qwen2.5:32b      # ~22 GB RAM
ollama pull llama3.3:70b     # ~45 GB, M4 Max/Ultra-Gebiet

Dann in Einstellungen → Profile → Neues Profil: Ollama / http://localhost:11434 / das geladene Modell wählen.

Migration

Wenn du ein v1.3.0-Apple-Intelligence-Profil gespeichert hattest, bleibt es in UserDefaults, taucht aber nicht mehr im Provider-Picker auf. Manuell unter Einstellungen → Profile löschen. Sonst keine Migration nötig.

Installation

  1. blitzbot-1.3.2-macos-arm64.zip herunterladen
  2. Entpacken, blitzbot.app nach /Programme ziehen
  3. Erster Start: Rechtsklick → Öffnen
  4. macOS 15+ meldet „beschädigt"? xattr -dr com.apple.quarantine /Applications/blitzbot.app