Qwen3.5-35B-A3B-IQ3_S-MTP

Qwen3.5-35B-A3B — IQ3_S with MTP Tensors

One of the first GGUF quantizations of Qwen3.5-35B-A3B that preserves the Multi-Token Prediction (MTP) head. Most public quantizations strip MTP tensors during conversion. This GGUF retains them, enabling speculative decoding via MTP.

Key specs

MTP Results

Tested with ik_llama.cpp (patched for Qwen3.5 MoE MTP support):

Why MTP is slower here

The MTP layer (layer 40) is itself a full MoE layer with 256 experts. On consumer hardware with CPU expert offloading (-ncmoe 4), each MTP forward costs nearly as much as a main model forward. The bottleneck is MoE expert transfer over PCIe, not the number of forward passes.

This is an informative negative result: MTP speculative decoding is not cost-effective on MoE models where the MTP head is also MoE. It would work on models with a dense MTP head.

MTP patches for ik_llama.cpp

This GGUF requires patched ik_llama.cpp with MTP support for Qwen3.5 MoE. The 5 patches (fixing 8 bugs in tensor mapping, compute buffers, and MoE handling) are available at: chimere/patches/ik-llama-mtp

How it was built

# Inject MTP tensors from original BF16 shards into IQ3_S base
python build_mtp_gguf_v3.py
# Source: IQ3_S custom-mix base + shards 13-14 from Qwen/Qwen3.5-35B-A3B

Build tool: ramp-quant/tools/build_mtp_gguf_v3.py

Related

• chimere — Rust inference runtime + MTP patches
• Qwen3.5-35B-A3B-RAMP-v2-15G — Optimized quantization without MTP

Author

Kevin Remondiere — Independent ML researcher, Oloron-Sainte-Marie, France

License

Apache 2.0. Base model follows Qwen's license.