Qwen 3.6 Model Configs

llama-server configurations optimized for coding on Vulkan.

Qwen3.6-27B MTP

27B parameter IQ4 model with Multi-Token Prediction (MTP) for speculative decoding. Uses a Q8 nextn draft model for fast token generation. Best all-round config for coding. Running on a single 16GB AMD 6800

new_27B_mtp.sh — Normal — optimized for coding, 32K context headless
new_27B_mtp.sh_think — Think — reasoning mode with higher temperature
new_27B_mtp.sh_chat — Chat — more loosey for conversational use

Qwen3.6-27B NGRAM

IQ4_XS quantized 27B model with light ngram-based speculative decoding intead of MTD. Running on a single 16GB AMD 6800 It consumes less VRAM and compute than the MTD version for 1/3 of prediction performances, more context available, better suited for domains where prediction don't matter much.

new_27B_xxs.sh — Normal — ngram-mod speculative decoding, coding optimized
new_27B_xxs.sh_think — Think — higher temperature for reasoning

Qwopus3.5-9B Coder

9B coder model (Qwopus3.5) with MTP speculative decoding. Smaller and faster, good for quick coding tasks yet slower than MoE.

new_9b.sh — Normal — 81K context, Q6_K quantization
new_9b.sh_think — Think — 30K context, higher temperature

Qwen3.6-35B-A3B MoE

Mixture-of-Experts model with 35B total / 3B active parameters. Running on a single 16GB AMD 6800Multiple configs ranging from fast IQ3_M quant to larger Q4_K_XL.

qwen_MoE.sh_ngram — Normal — IQ3_M quant, ngram speculative decoding
qwen_MoE.sh_ngram_think — Think enabled NGRAM — 121K context
fastest_MoE.sh_mtd — Fastest MTD with MTP draft all loaded in VRAM
moe_mtd_xl.sh — Large MTD XL — Q4_K_XL quant, 80K fit-ctx

Qwen3.6-35B-A3B Base

Base configuration for the 35B-A3B UD model with Q4_K_S quantization, no MTD heads. Good reference config with 80K context. Running on a single 16GB AMD 6800

qwen3.6.sh — Base — Q4_K_S UD, ngram speculative decoding

Generated from lm/ scripts — 5 model families, 12 configs total.