@UnslothAI just dropped MTP for Qwen 3.7 Flash Next (125B A6B)! 25 tokens/sec on a single RTX 4090! I took the 125B (6B active) setup from below, plugged in the new shared-Q8_0 MTP drafter, and pushed decode throughput to 25.35 tokens/sec at an 80k context window on a single 24GB RTX 4090. You must build from @danielhanchen's PR branch (#144): git clone github.com/unslothai/llama.c… && cd llama.cpp git fetch origin pull/144/head:mtp && git checkout mtp cmake -B build -DCMAKE_BUILD_TYPE=Release -DGGML_CUDA=ON -DCMAKE_CUDA_ARCHITECTURES=native cmake --build build -j $(nproc) --target llama-server # The 80k MTP Benchmark Run Fed it the same continuous 28k token prompt, combining CPU-MoE offload (-cmoe) with MTP speculative decoding (--spec-draft-n-max 2): llama.cpp flags: ./build/bin/llama-server -m Qwen3.8-Flash-Next-UD-Q4_K_XL-00001-of-00004.gguf --spec-type draft-mtp --spec-draft-model mtp-Qwen3.8-Flash-Next-shared-Q8_0.gguf --spec-draft-n-max 2 -c 80000 -cmoe --threads 12 -b 4096 -ub 4096 --port 8080 Prefill: 471.08 t/s (Up from 355.72 t/s baseline, a 32% prefill jump) Decode: 25.35 t/s (Up from 20.84 t/s baseline) Footprint: 19.0 GB VRAM | 92 GB DDR4 System RAM # Why shared-Q8_0 is the Optimal Choice The shared-Q8_0 MTP head borrows token embeddings and output projections directly from the base model, saving ~1.3GB of critical VRAM. Setting --spec-draft-n-max 2 delivers consistent speculative acceptance without thrashing the PCIe 4.0 bus during offloaded MoE passes. This is just the initial 80k verification pass. Currently benchmarking the 180,000 and 250,000 deep context runs with MTP active. full scaling data drops soon. Hugging Face GGUF and MTP draft links are in the replies below (available in same Unsloth's repo). Have you tested speculative decoding on Qwen 3.8 Flash Next yet? Lemme know if u have any questions