ProxMenux/AppImage/config/verified_ai_models.json

{
  "_description": "Verified AI models for ProxMenux notifications. Only models listed here will be shown to users. Models are tested to work with the chat/completions API format.",
  "_updated": "2026-09-02",
  "_verifier": "Refreshed by .github/workflows/verify-ai-models.yml (daily). The workflow runs .github/scripts/ai-models-verifier/verify.py against every provider whose API key is configured in repository Secrets, then applies the report via apply.py — which honours per-provider `_exclude` lists. `_exclude` is intentionally minimal: it only drops models that are technically incapable of generating a chat completion for a normal prompt (safety classifiers, agentic-only endpoints, meta-routers, wrong modalities). Everything else the verifier passes is surfaced — including language-specialised models (Arabic, Chinese, ...), reasoning models, legacy families and dated snapshots — so a user with a specific need can still pick the model that fits. Manually re-run from the Actions tab (any branch) when a new model needs to be picked up out of cycle.",

  "groq": {
    "models": [
      "llama-3.3-70b-versatile",
      "llama-3.1-8b-instant",
      "meta-llama/llama-4-scout-17b-16e-instruct",
      "openai/gpt-oss-120b",
      "openai/gpt-oss-20b"
    ],
    "recommended": "llama-3.3-70b-versatile",
    "_exclude": [
      "openai/gpt-oss-safeguard-*",
      "groq/compound",
      "groq/compound-*"
    ],
    "_note": "`_exclude` covers models the verifier may technically pass but that do not produce a usable chat completion: openai/gpt-oss-safeguard-* is a safety classifier (returns a category, not free text); groq/compound* is an agentic system that expects multi-step tool use, not a plain prompt."
  },

  "gemini": {
    "models": [
      "gemini-flash-lite-latest",
      "gemini-2.5-flash-lite",
      "gemini-2.5-flash",
      "gemini-3.1-flash-lite",
      "gemini-3-flash-preview",
      "gemini-3.5-flash"
    ],
    "recommended": "gemini-2.5-flash-lite",
    "_exclude": [
      "gemini-embedding-*",
      "gemini-*-pro*",
      "gemini-*-thinking*"
    ],
    "_deprecated": ["gemini-2.0-flash", "gemini-2.0-flash-lite", "gemini-1.5-flash", "gemini-1.0-pro", "gemini-pro"],
    "_note": "`_exclude` drops embeddings (wrong modality) and Pro / thinking variants that reject `thinkingConfig.thinkingBudget: 0` and therefore never return a visible completion within a reasonable token budget — technical failure with our current provider config."
  },

  "openai": {
    "models": [
      "gpt-4.1-nano",
      "gpt-4.1-mini",
      "gpt-4o-mini",
      "gpt-4.1",
      "gpt-4o",
      "gpt-5-chat-latest",
      "gpt-5-nano"
    ],
    "recommended": "gpt-4.1-nano",
    "_exclude": [
      "gpt-4o-audio*",
      "gpt-4o-realtime*",
      "gpt-4o-search*",
      "gpt-4o-transcribe*",
      "gpt-4o-mini-audio*",
      "gpt-4o-mini-realtime*",
      "gpt-4o-mini-search*",
      "gpt-4o-mini-transcribe*",
      "gpt-4o-mini-tts",
      "computer-use-*"
    ],
    "_note": "`_exclude` covers wrong-modality variants (audio, realtime, search, transcribe, tts) that cannot handle a plain notification-translation prompt, plus computer-use which requires an agent loop. All other OpenAI chat/completion models are surfaced — including legacy families (gpt-3.5, gpt-4), reasoning models (o-series, gpt-5.x non-chat) and dated snapshots — so users can pick by their own criteria (cost, quality, reproducibility). openai_provider.py already handles reasoning models via max_completion_tokens + reasoning_effort=minimal."
  },

  "anthropic": {
    "models": [
      "claude-haiku-4-5",
      "claude-sonnet-5",
      "claude-opus-4-8",
      "claude-sonnet-4-6",
      "claude-opus-4-6",
      "claude-fable-5"
    ],
    "recommended": "claude-haiku-4-5",
    "_exclude": [],
    "_note": "No technical exclusions — every Claude generation returns free-text completions for a plain prompt. The verifier tests each model listed under `models`; if a specific ID stops working upstream it simply drops out of the passing set. Anthropic does not expose a public models-list API, so new models must be added to `models` manually before the verifier can test them."
  },

  "openrouter": {
    "models": [
      "meta-llama/llama-3.3-70b-instruct",
      "meta-llama/llama-3.1-70b-instruct",
      "meta-llama/llama-3.1-8b-instruct",
      "meta-llama/llama-4-scout",
      "anthropic/claude-haiku-4.5",
      "anthropic/claude-sonnet-4.6",
      "google/gemini-2.5-flash-lite",
      "google/gemini-2.5-flash",
      "openai/gpt-4o-mini",
      "mistralai/mistral-small-3.2-24b-instruct",
      "nvidia/nemotron-3-super-120b-a12b:free",
      "google/gemma-4-26b-a4b-it:free",
      "nvidia/nemotron-nano-12b-v2-vl:free",
      "nvidia/nemotron-3-nano-30b-a3b:free",
      "poolside/laguna-s-2.1:free",
      "nvidia/nemotron-3-nano-omni-30b-a3b-reasoning:free",
      "openai/gpt-oss-20b:free"
    ],
    "recommended": "meta-llama/llama-3.3-70b-instruct",
    "_exclude": [
      "openrouter/*",
      "*/*-audio*",
      "*/*-audio-*",
      "*/*-tts*",
      "*/*-whisper*",
      "*/*-embed*",
      "*/*-embedding*",
      "*/*-image*",
      "*/*-vision-only*"
    ],
    "_note": "OpenRouter aggregates hundreds of models; the free-tier variants (:free suffix) are intentionally supported per user request and never blocked. `_exclude` covers only meta-routers (`openrouter/free`, `openrouter/auto` — they route dynamically to something else, so their behaviour is not the model the user picked) and wrong-modality models (audio, tts, whisper, embeddings, image, vision-only). Everything else — chat models across every family, language and price tier — is surfaced so the user can pick the fit."
  },

  "ollama": {
    "_note": "Ollama models are local, we don't filter them. User manages their own models.",
    "models": [],
    "recommended": ""
  }
}