Qwen3.6-35B-A3B-EXTENSOR
An EXTENSOR runtime image derived from Qwen/Qwen3.6-35B-A3B. It contains the text target and its MTP draft module; the vision tower is not included. EXTENSOR uses MTP for lossless greedy speculative decoding while the target remains the authority for every emitted token.
Artifact
- File:
Qwen3.6-35B-A3B-EXTENSOR-ROCmFP4-MTP-v2.extensor.gguf - Size: 19,659,370,496 bytes
- SHA-256:
4844f6145ece96b433b494cf87fd2da700b87333508edccbb78cfba53f953de1 - Quantization: ROCmFP4/ROCmFP4_FAST target and MTP, Q8_0 MTP fusion, F16 controls
- Format: GGUF v3, EXTENSOR pack v2, qmoe-v4 target slab, resident MTP expert bank
Requires EXTENSOR 2.4.0 or a newer build with backend API 4. Exact MTP verification requires the complete target expert slab resident. This file is not compatible with llama.cpp.
The target-only v1 artifact remains supported by the separate
qwen35moe-rocmfp4-runtime-v1 backend profile:
- File:
Qwen3.6-35B-A3B-EXTENSOR-ROCmFP4-v1.extensor.gguf - Size: 19,205,345,280 bytes
- SHA-256:
66260013d3ec7b880b9d148fc688a2b17f7dafcdadcdc3599738db7302edf896
Source
Qwen/Qwen3.6-35B-A3B at revision
995ad96eacd98c81ed38be0c5b274b04031597b0.
License
Apache License 2.0. See the upstream repository for usage and citation details.