{
  "version": 1,
  "generatedAt": "2026-08-10T01:00:00Z",
  "schema": {
    "engineModelType": ["gemma_it", "gemma_pt", "litertlm", "llama_cpp"],
    "engineFileType": ["litertlm", "task", "gguf", "tflite"],
    "tier": ["lite", "balanced", "advanced"],
    "license": ["apache-2.0", "mit", "gemma", "llama3.1", "other"]
  },
  "note": "Hosted at https://dharmica.trivartha.com/ai/manifest.json. The Android app fetches this on launch and every refresh. Edit this file to add a new model, bump a version, or pin a HuggingFace revision — no app update, no store release required. Every entry below has been verified against the live HuggingFace tree API; sizes match the real on-Hub content-length.",
  "models": [
    {
      "id": "qwen3-0_6b@1.0.0",
      "family": "qwen",
      "name": "Qwen3 0.6B",
      "version": "1.0.0",
      "tier": "lite",
      "license": "apache-2.0",
      "publisher": "Alibaba (LiteRT build)",
      "repoId": "litert-community/Qwen3-0.6B",
      "filePattern": "qwen3_0_6b_mixed_int4\\.litertlm$",
      "pinnedRevision": "main",
      "downloadBytes": 497664000,
      "minRamBytes": 3221225472,
      "paramsLabel": "0.6B",
      "quantLabel": "mixed int4",
      "contextTokens": 4096,
      "multilingual": true,
      "summary": "Smallest usable option and the fastest to download. Handles short summaries and follow-up drafts on entry-level phones.",
      "releaseNotes": "Initial release. Recommended for 3 GB phones.",
      "engineModelType": "litertlm",
      "engineFileType": "litertlm"
    },
    {
      "id": "qwen2.5-0.5b-it@1.0.0",
      "family": "qwen",
      "name": "Qwen2.5 0.5B Instruct",
      "version": "1.0.0",
      "tier": "lite",
      "license": "apache-2.0",
      "publisher": "Alibaba (LiteRT build)",
      "repoId": "litert-community/Qwen2.5-0.5B-Instruct",
      "filePattern": "Qwen2\\.5-0\\.5B-Instruct_multi-prefill-seq_q8_ekv1280\\.task$",
      "pinnedRevision": "main",
      "downloadBytes": 546660344,
      "minRamBytes": 3221225472,
      "paramsLabel": "0.5B",
      "quantLabel": "int8",
      "contextTokens": 1280,
      "multilingual": true,
      "summary": "Tiny but capable. Cheapest option for 3 GB phones; keep prompts short.",
      "releaseNotes": "Initial release.",
      "engineModelType": "gemma_it",
      "engineFileType": "task"
    },
    {
      "id": "qwen3.5-0.8b@1.0.0",
      "family": "qwen",
      "name": "Qwen3.5 0.8B",
      "version": "1.0.0",
      "tier": "lite",
      "license": "apache-2.0",
      "publisher": "Alibaba (LiteRT build)",
      "repoId": "litert-community/Qwen3.5-0.8B",
      "filePattern": "Qwen3\\.5-0\\.8B_int8\\.litertlm$",
      "pinnedRevision": "main",
      "downloadBytes": 978249216,
      "minRamBytes": 3221225472,
      "paramsLabel": "0.8B",
      "quantLabel": "int8",
      "contextTokens": 4096,
      "multilingual": true,
      "summary": "Newer Qwen 3.5 generation at 0.8B — better drafting than the 0.6B at a similar download size.",
      "releaseNotes": "Initial release.",
      "engineModelType": "litertlm",
      "engineFileType": "litertlm"
    },
    {
      "id": "qwen2.5-1.5b-it@1.0.0",
      "family": "qwen",
      "name": "Qwen2.5 1.5B Instruct",
      "version": "1.0.0",
      "tier": "balanced",
      "license": "apache-2.0",
      "publisher": "Alibaba (LiteRT build)",
      "repoId": "litert-community/Qwen2.5-1.5B-Instruct",
      "filePattern": "Qwen2\\.5-1\\.5B-Instruct_multi-prefill-seq_q8_ekv4096\\.litertlm$",
      "pinnedRevision": "main",
      "downloadBytes": 1597931520,
      "minRamBytes": 4294967296,
      "paramsLabel": "1.5B",
      "quantLabel": "int8",
      "contextTokens": 4096,
      "multilingual": true,
      "summary": "Solid all-rounder for guidance drafting. Good Hindi and regional-language handling.",
      "releaseNotes": "Initial release. Default pick for 4 GB+ devices.",
      "engineModelType": "gemma_it",
      "engineFileType": "litertlm"
    },
    {
      "id": "qwen3-1_7b@1.0.0",
      "family": "qwen",
      "name": "Qwen3 1.7B",
      "version": "1.0.0",
      "tier": "balanced",
      "license": "apache-2.0",
      "publisher": "Alibaba (LiteRT build)",
      "repoId": "litert-community/Qwen3-1.7B",
      "filePattern": "Qwen3_1\\.7B\\.litertlm$",
      "pinnedRevision": "main",
      "downloadBytes": 2056729520,
      "minRamBytes": 6442450944,
      "paramsLabel": "1.7B",
      "quantLabel": "int8",
      "contextTokens": 4096,
      "multilingual": true,
      "summary": "Noticeably better written output than the 1.5B. Worth the extra download on a 6 GB phone or better.",
      "releaseNotes": "Initial release. Recommended for 6 GB+ devices.",
      "engineModelType": "litertlm",
      "engineFileType": "litertlm"
    },
    {
      "id": "gemma3-1b-it@1.0.0",
      "family": "gemma",
      "name": "Gemma 3 1B IT",
      "version": "1.0.0",
      "tier": "balanced",
      "license": "apache-2.0",
      "publisher": "Google (LiteRT build)",
      "repoId": "litert-community/Gemma3-1B-IT",
      "filePattern": "Gemma3-1B-IT_multi-prefill-seq_q4_block128_ekv4096\\.task$",
      "pinnedRevision": "main",
      "downloadBytes": 689308662,
      "minRamBytes": 4294967296,
      "paramsLabel": "1B",
      "quantLabel": "int4",
      "contextTokens": 4096,
      "multilingual": true,
      "summary": "Smallest Gemma 3, ungated. Tight little draft replies for 4 GB phones.",
      "releaseNotes": "Initial release.",
      "engineModelType": "gemma_it",
      "engineFileType": "task"
    },
    {
      "id": "deepseek-r1-distill-qwen-1.5b@1.0.0",
      "family": "qwen",
      "name": "DeepSeek R1 Distill Qwen 1.5B",
      "version": "1.0.0",
      "tier": "balanced",
      "license": "mit",
      "publisher": "DeepSeek (LiteRT build)",
      "repoId": "litert-community/DeepSeek-R1-Distill-Qwen-1.5B",
      "filePattern": "DeepSeek-R1-Distill-Qwen-1\\.5B_multi-prefill-seq_q8_ekv4096\\.litertlm$",
      "pinnedRevision": "main",
      "downloadBytes": 1833451520,
      "minRamBytes": 6442450944,
      "paramsLabel": "1.5B",
      "quantLabel": "int8",
      "contextTokens": 4096,
      "multilingual": true,
      "summary": "Reasoning-tuned distill. Better at structured verse analysis than base Qwen 1.5B; slower first-token.",
      "releaseNotes": "Initial release. MIT licensed.",
      "engineModelType": "gemma_it",
      "engineFileType": "litertlm"
    },
    {
      "id": "phi-4-mini-it@1.0.0",
      "family": "phi",
      "name": "Phi-4 Mini Instruct",
      "version": "1.0.0",
      "tier": "balanced",
      "license": "mit",
      "publisher": "Microsoft Research",
      "repoId": "litert-community/Phi-4-mini-instruct",
      "filePattern": "Phi-4-mini-instruct_multi-prefill-seq_q8_ekv4096\\.litertlm$",
      "pinnedRevision": "main",
      "downloadBytes": 3910090752,
      "minRamBytes": 6442450944,
      "paramsLabel": "3.8B",
      "quantLabel": "int8",
      "contextTokens": 4096,
      "multilingual": true,
      "summary": "Strong reasoning for its size. Slow first-token on mid-range silicon.",
      "releaseNotes": "Initial release. MIT licensed, ungated.",
      "engineModelType": "litertlm",
      "engineFileType": "litertlm"
    },
    {
      "id": "gemma-4-e2b-it@1.0.0",
      "family": "gemma",
      "name": "Gemma 4 E2B Instruct",
      "version": "1.0.0",
      "tier": "advanced",
      "license": "apache-2.0",
      "publisher": "Google (LiteRT build)",
      "repoId": "litert-community/gemma-4-E2B-it-litert-lm",
      "filePattern": "gemma-4-E2B-it\\.litertlm$",
      "pinnedRevision": "main",
      "downloadBytes": 2588147712,
      "minRamBytes": 8589934592,
      "paramsLabel": "2B effective",
      "quantLabel": "int4",
      "contextTokens": 4096,
      "multilingual": true,
      "summary": "Best drafting quality of the set, flagship phones only. Apache-2.0 and ungated.",
      "releaseNotes": "Initial release. Recommended for 8 GB+ devices.",
      "engineModelType": "gemma_it",
      "engineFileType": "litertlm"
    },
    {
      "id": "gemma3-4b-it@1.0.0",
      "family": "gemma",
      "name": "Gemma 3 4B IT",
      "version": "1.0.0",
      "tier": "advanced",
      "license": "apache-2.0",
      "publisher": "Google (LiteRT build)",
      "repoId": "litert-community/Gemma3-4B-IT",
      "filePattern": "gemma3-4b-it-int4-web\\.task$",
      "pinnedRevision": "main",
      "downloadBytes": 2559442944,
      "minRamBytes": 8589934592,
      "paramsLabel": "4B",
      "quantLabel": "int4",
      "contextTokens": 4096,
      "multilingual": true,
      "summary": "Stronger drafting than the 1B, still Apache-2.0 and ungated. For 8 GB+ phones.",
      "releaseNotes": "Initial release.",
      "engineModelType": "gemma_it",
      "engineFileType": "task"
    },
    {
      "id": "qwen3-8b@1.0.0",
      "family": "qwen",
      "name": "Qwen3 8B (mixed int4)",
      "version": "1.0.0",
      "tier": "advanced",
      "license": "apache-2.0",
      "publisher": "Alibaba (LiteRT build)",
      "repoId": "litert-community/Qwen3-8B",
      "filePattern": "qwen3_8b_mixed_int4\\.litertlm$",
      "pinnedRevision": "main",
      "downloadBytes": 4887412736,
      "minRamBytes": 10737418240,
      "paramsLabel": "8B",
      "quantLabel": "mixed int4",
      "contextTokens": 4096,
      "multilingual": true,
      "summary": "Largest option. Long first-token latency on mid-range silicon; best quality on flagships.",
      "releaseNotes": "Initial release. For 10 GB+ devices.",
      "engineModelType": "litertlm",
      "engineFileType": "litertlm"
    }
  ]
}