{
  "schema_version": 1,
  "comment": "Pre-trained model catalog served at https://repo.eldric.ai/catalog/models.json. Customer-facing index. Schema v1 from OPS Wave 10 ship 74. Wave 21 ship: nomic-embed-text-v1.5 GGUF added as a customer-self-host fallback for the RAG embedding backend. Primary GA backend is Ollama-served nomic-embed-text on inference nodes (already cluster-resident); the GGUF here is for customers running offline / air-gapped / non-Ollama topologies.",
  "updated_at": "2026-05-19T16:55Z",
  "entries_count": 4,
  "entries": [
    {
      "id": "nomic-embed-text-v1.5-Q4_K_M",
      "family": "nomic-embed-text",
      "version": "v1.5-Q4_K_M",
      "type": "gguf",
      "filename": "nomic-embed-text-v1.5.Q4_K_M.gguf",
      "size_mb": 81,
      "sha256": "d4e388894e09cf3816e8b0896d81d265b55e7a9fff9ab03fe8bf4ef5e11295ac",
      "license": "apache-2.0",
      "description": "Nomic AI's nomic-embed-text v1.5 embedding model, Q4_K_M quantization (~80 MB). 768-dim output; multilingual (EN + DE + many others); long-context (8192 tokens). Customer-self-host fallback for offline / air-gapped / non-Ollama installs. Primary v5.0 GA topology serves this same model via Ollama on inference nodes; cluster-resident installs do not need this download.",
      "upstream": "https://huggingface.co/nomic-ai/nomic-embed-text-v1.5-GGUF",
      "embedding_dim": 768,
      "max_context_tokens": 8192,
      "download_url": "https://repo.eldric.ai/models/nomic-embed-text-v1.5.Q4_K_M.gguf",
      "role": "embedding-fallback",
      "primary_backend": "Ollama (nomic-embed-text) on inference nodes",
      "minimum_eldric_version": "5.0.0",
      "wave": "OPS Wave 21 (2026-05-19)"
    },
    {
      "id": "eldric-router-v18-gqa",
      "family": "eldric-router",
      "version": "v18-gqa",
      "type": "enrn",
      "filename": "eldric-router-v18-gqa.enrn",
      "size_mb": 4,
      "sha256": "f3cf3bbbeb8d82834fba53b8c6df08f0a6516ee904595b22c71fac03db3fb497",
      "license": "proprietary",
      "description": "xLSTM intent-routing model, v18 with GQA (Grouped-Query Attention). Trained on real intent corpus by TECH Wave 17/18 GQA student pipeline. Retired as canonical 2026-05-19 — see status field. File remains served for any customer who already pinned this version; new installs should use the LLM-only routing fallback.",
      "status": "retired-canonical — research-grade only; superseded by router v19/v20 training plan",
      "accuracy_measured": 0.58,
      "accuracy_target": 0.90,
      "retirement_reason": "Below accuracy gate. Single-step distillation structural cap; v19/v20 cascade is the structurally-correct fix.",
      "retirement_date": "2026-05-19",
      "supersession_spec": "docs/specs/2026-05-19-router-v19-v20-training-data-spec.md",
      "intents": [
        "PlainChat", "RAGQuery", "AgentInvoke", "ScienceQuery", "MediaRequest",
        "CommRequest", "TrainingRequest", "MemoryStore", "MemoryRecall",
        "DataOperation", "IoTRequest", "ADMIN", "SwarmRequest"
      ],
      "vocab_filename": "router-xlstm-v18-tok-vocab.json",
      "vocab_sha256": "b05fce4f881bd1b94cbe6c89ea9896793a59b23b1b615e3057dea934cd270825",
      "vocab_url": "https://repo.eldric.ai/models/router-xlstm-v18-tok-vocab.json",
      "download_url": "https://repo.eldric.ai/models/eldric-router-v18-gqa.enrn",
      "minimum_eldric_version": "5.0.0-292.alpha118ppp",
      "trained_on": "TECH Wave 17/18 real GQA student pipeline (alpha118ttt cadence)"
    },
    {
      "id": "eldric-router-v17",
      "family": "eldric-router",
      "version": "v17",
      "type": "enrn",
      "filename": "eldric-router-v17.enrn",
      "size_mb": null,
      "sha256": null,
      "license": "proprietary",
      "description": "PLACEHOLDER — Wave 11 ship 74 schema-validation file. Retained as a pinnable historic entry. Use eldric-router-v18-gqa.enrn for real intent routing.",
      "intents": [
        "PlainChat", "RAGQuery", "AgentInvoke", "ScienceQuery", "MediaRequest",
        "CommRequest", "TrainingRequest", "MemoryStore", "MemoryRecall",
        "DataOperation", "IoTRequest", "ADMIN", "SwarmRequest"
      ],
      "vocab_filename": "router-xlstm-v17-tok-vocab.json",
      "vocab_url": "https://repo.eldric.ai/models/router-xlstm-v17-tok-vocab.json",
      "download_url": "https://repo.eldric.ai/models/eldric-router-v17.enrn",
      "minimum_eldric_version": "5.0.0-220.alpha118fff",
      "status": "placeholder — superseded by v18-gqa"
    },
    {
      "_comment_template": "Example EMM entry — distilled corpus checkpoint. Still placeholder; waits on TECH distillation pipeline output.",
      "id": "eldric-papers-at-emm-v1",
      "family": "eldric-papers-at-emm",
      "version": "v1",
      "type": "emm",
      "filename": "eldric-papers-at-emm-v1.emm",
      "size_mb": null,
      "sha256": null,
      "license": "proprietary",
      "description": "Pre-distilled Eldric Matrix Memory checkpoint for the ai-papers-at corpus (Austrian AI research, oamonitor-enriched). Drops into /data/eldric/memory/. Substitute for re-ingesting 167k chunks against your own embeddings.",
      "intents": [],
      "download_url": "https://repo.eldric.ai/models/eldric-papers-at-emm-v1.emm",
      "latest_alias": "eldric-papers-at-emm-latest.emm",
      "minimum_eldric_version": "5.0.0-220.alpha118fff",
      "status": "placeholder — TECH distillation pipeline not yet shipped"
    }
  ]
}