# Pinned model inputs

Use the **Pinned model acquisition** command in the chart row, or the same command in the [current command index](../portable/commands.json). It verifies and extracts the exact launch package before calling its acquisition tool. Follow the [portable launch guide](../PORTABLE-LAUNCH.md) for host requirements.

Choose your own persistent model directory. Current Qwen3.8 commands use `MODEL_DIR` for the checkpoint directory itself. Historical Qwen3.6 commands use `MODEL_DIR` for a collection directory containing the required `qwen36-dense`, `qwen36-moe` or `qwen36-gguf` inputs. The original laboratory directories are not required. The launcher resolves the chosen host directory and mounts the verified input read-only at the image's expected internal path.

The acquisition tools reuse matching files after hash verification. Existing mismatches and interrupted partial downloads are preserved and cause an error. A missing payload requires enough free space for that payload and the tool's reserve. The current end-to-end campaign reuses already verified full checkpoints, as authorized; it does not claim to have downloaded every model again.

| Input | Pinned public revision | Manifest bytes |
|---|---|---:|
| [Qwen/Qwen3.8-27B](https://huggingface.co/Qwen/Qwen3.8-27B/tree/1d4bf0f2ff6012fd82039f2fa52739d0dd7c60c0) | `1d4bf0f2ff6012fd82039f2fa52739d0dd7c60c0` | 55,586,114,863 |
| [Qwen/Qwen3.6-27B](https://huggingface.co/Qwen/Qwen3.6-27B/tree/6a9e13bd6fc8f0983b9b99948120bc37f49c13e9) | `6a9e13bd6fc8f0983b9b99948120bc37f49c13e9` | 55,586,107,940 |
| [Qwen/Qwen3.6-35B-A3B](https://huggingface.co/Qwen/Qwen3.6-35B-A3B/tree/995ad96eacd98c81ed38be0c5b274b04031597b0) | `995ad96eacd98c81ed38be0c5b274b04031597b0` | 71,926,865,825 |

Qwen3.8 FP16, local RTN8 and local RTN4 configurations all start from the same full checkpoint. Quantized runtimes create their packed RTN weights during loading. Neither a separate GPTQ/AWQ download nor model weights embedded in the runtime image are required.

The historical dense F16 GGUF command additionally acquires the 54,657,732,704-byte dense F16 GGUF and its pinned text configuration. It validates all 28 parts and the assembled SHA-256 `b2347376b9bb7d12cf5f1d31c53ac4c60dd3d4b95068a09351187b328b5e9d89`. Its dense HF input supplies the pinned tokenizer. It does not acquire the unused MoE GGUF. Peak new disk use includes the assembled file and one part, rather than a second complete set of parts.

Public input availability and identities are documented in [the URL verification](../public-model-url-verification.json). Runtime and model identities are separate. Historical reproduction disables network model lookup and pins the tokenizer cache reference. Earlier original GGUF commands omitted `--tokenizer-revision`; their complete original tokenizer cache identity was not recovered, so that historical limitation remains explicit.

