---

license: apache-2.0

base_model: Qwen/Qwen3.5-122B-A10B

tags:

• gguf
• qwen3.5
• enclave

---

qwen3.5-122b-gguf (Enclave model volume)

Curated model volume for Enclave confidential inference:

Qwen3.5-122B-A10B at unsloth Dynamic IQ4_XS (~60 GB), **merged into a single

GGUF** and bundled with the official tokenizer.json so the volume is

self-contained.

Upstream ships this quantization as a 3-part split family (classic Git-LFS

repos cap files at 50 GB). Enclave's pinned wasmtime ggml toolchain predates

split-family loading but serves single-file GGUFs, so this repo carries the

merge instead (Xet storage, no 50 GB cap). The merge was produced with

llama.cpp's llama-gguf-split --merge (build bec4772): tensor data is a

byte-for-byte copy of upstream — only the GGUF header bookkeeping changes

(split.count → 0, 879 tensors in one file).

Provenance

| File | Upstream | Revision | sha256 |

|------|----------|----------|--------|

| Qwen3.5-122B-A10B-UD-IQ4_XS.gguf | merged from unsloth/Qwen3.5-122B-A10B-GGUF UD-IQ4_XS/ | 51eab4d59d53f573fb9206cb3ce613f1d0aa392b | 9c9701c1673f80cc164bf66e3a82b957dc18a7223676339a9bbd76175c4d8f92 |

| tokenizer.json | Qwen/Qwen3.5-122B-A10B | dc4d348443bc740c68e2d77492492c11606384d5 | 5f9e4d4901a92b997e463c1f46055088b6cca5ca61a6522d1b9f64c4bb81cb42 |

Merge inputs — the upstream split parts, byte-identical both in

unsloth's repo and in this repo's prior revision 8beb6427:

| Split part | sha256 |

|------------|--------|

| Qwen3.5-122B-A10B-UD-IQ4_XS-00001-of-00003.gguf | 1bd4c860d2300566087683d8b971830eaea466eec8f8c5ee69f9ddc2bcea3613 |

| Qwen3.5-122B-A10B-UD-IQ4_XS-00002-of-00003.gguf | 1f77e62d600cb255e327e76ffa058755d1de2b8a1872dd58c89eeb5f8c90e7e8 |

| Qwen3.5-122B-A10B-UD-IQ4_XS-00003-of-00003.gguf | c644a84cedb93cb85aaedc9bf711d6a388d8205a59566638a9eb1f5c8afd04f2 |

LICENSE is the Apache-2.0 text; both upstreams are Apache-2.0.

Usage on Enclave

Wrapped as a Tinfoil Modelwrap volume (dm-verity; the root hash is part of the

enclave measurement). Deployments attach it by name and the guest reads it at

/models/<name>; the host preloads the GGUF as the wasi-nn ggml graph. As a

single-file GGUF it loads on wasmtime toolchains without split-family support,

and with the tokenizer bundled, llm-chat's default tokenizer.json lookup

works with no cross-volume configuration.