qwen3.5-0.8b-gguf

qwen3.5-0.8b-gguf (Enclave model volume)

Curated model volume for Enclave confidential inference: Qwen3.5-0.8B at Q8_0 (~775 MB), a single-file GGUF bundled with the official tokenizer.json so the volume is self-contained.

This is the family's SPECULATIVE-DECODING DRAFT volume: attach it next to qwen3.5-9b-gguf, qwen3.6-27b-gguf, or qwen3.5-122b-gguf and name it in the llm-chat config ("draft": "qwen3.5-0.8b-gguf") to have it propose tokens the big model verifies in one batched pass. Output is byte-for-byte the big model's own (exact-match verification); the draft only changes speed. Q8_0 rather than a 4-bit quant because a draft's sole job is predicting the target's picks - the ~0.4 GB saved by 4-bit is not worth the acceptance-rate loss.

Upstream unsloth/Qwen3.5-0.8B-GGUF ships five quantizations plus a vision mmproj and no tokenizer.json. This volume carries ONLY the one served quant plus the tokenizer (same rationale as the sibling volumes): a multi-gguf volume makes the host's preload pick ambiguous, and the extra files never serve.

Provenance

The tokenizer is byte-identical to the ones in qwen3.5-9b-gguf, qwen3.6-27b-gguf and qwen3.5-122b-gguf (the Qwen3.5/3.6 family shares it - same sha256, vocab 248320, <|im_end|> 248046 / <|endoftext|> 248044), which is exactly the property the speculative-decoding compatibility gate requires.

Model geometry (GGUF arch qwen35, hybrid-SSM): 24 layers, full_attention_interval 4 (6 KV layers), 2 KV heads, head_dim 256.

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; the wasm-manager mounts it read-only at /models/qwen3.5-0.8b-gguf.