Apertus 70B Instruct

Provenance · Annex IV §1(c)

{
  "_doc": "EU AI Act Annex IV §1(c) — model lineage / supply-chain record.",
  "deployment_id": "ailiance/apertus-70b",
  "deployed_at_utc": "2026-04-22T00:00:00Z",
  "deployed_on_host": "studio (Mac Studio M3 Ultra)",
  "served_via": "uvicorn FastAPI worker @ studio:9301, MLX inference (Metal)",
  "source": {
    "provider": "Swiss AI Initiative (EPFL / ETHZ / CSCS)",
    "huggingface_repo": "swiss-ai/Apertus-70B-Instruct-2509",
    "huggingface_url": "https://huggingface.co/swiss-ai/Apertus-70B-Instruct-2509",
    "repo_commit_sha": "ec382cbace70",
    "repo_last_modified_utc": "2026-04-21T09:39:24Z",
    "license_spdx": "apache-2.0",
    "license_url": "https://huggingface.co/swiss-ai/Apertus-70B-Instruct-2509/blob/main/LICENSE"
  },
  "weights_file": {
    "filename": "(MLX 8-bit conversion in /Users/clems/ailiance-mac-tuner/models/Apertus-70B-Instruct-2509)",
    "size_bytes": null,
    "sha256_expected": null,
    "comment": "Original safetensors checksums on HF; MLX local conversion produced via mlx_lm.convert. Reconversion deterministic from upstream commit."
  },
  "architecture": {
    "family": "Apertus (Llama-style dense)",
    "total_parameters": 70599864480,
    "active_parameters_per_token": 70599864480,
    "attention": "standard MHA",
    "context_length_native": 65536
  },
  "quantization": {
    "method": "MLX 8-bit",
    "produced_by": "Ailiance (mlx_lm.convert from upstream BF16)"
  },
  "modifications_post_download": [
    "MLX 8-bit quantisation (mlx_lm.convert)",
    "Domain LoRA adapters loaded at runtime — see output/adapters/apertus/"
  ],
  "intended_use": "Public chat playground at https://ailiance.fr/chat/ailiance/apertus-70b — multilingual EU reasoning + electronics/hardware domains via LoRA adapters.",
  "out_of_scope": [
    "Healthcare diagnosis without clinician review",
    "Legal advice",
    "Autonomous safety-critical decisions"
  ]
}