Qwen3.6-35B-A3B — Strix Halo Optimised GGUFs

Dynamic mixed-precision GGUF quantizations of Qwen/Qwen3.6-35B-A3B, produced and benchmarked on a Framework Desktop with AMD Ryzen AI MAX+ 395 (Radeon 8060S, gfx1151, 128 GB UMA) running Vulkan via llama.cpp.

Variants

File	Size	prefill (t/s)	decode (t/s)	Notes
Qwen3.6-35B-A3B-Q8_0.gguf	35 GB	975	52.7	near-lossless reference
Qwen3.6-35B-A3B-Q6_K.gguf	27 GB	830	62.2
Qwen3.6-35B-A3B-Q5_K_M.gguf	24 GB	943	64.1
Qwen3.6-35B-A3B-Q4_K_M.gguf	20 GB	1021	70.2	production sweet spot
Qwen3.6-35B-A3B-Q4_0.gguf	19 GB	1061	76.5	fastest decode
Qwen3.6-35B-A3B-IQ4_NL.gguf	19 GB	891	73.1
Qwen3.6-35B-A3B-DYNAMIC.gguf	19 GB	1100	64.0	fastest prefill; mixed per-tensor quant

All numbers: pp=4096 tokens, tg=128 tokens; -fa 1 -ctk q8_0 -ctv q8_0 -ub 2048 -b 2048 on a single Vulkan gfx1151 device.

Dynamic mix recipe

DYNAMIC.gguf uses a per-tensor quantization map chosen for the hybrid Gated DeltaNet + Gated Attention architecture:

• attn_k / attn_q / attn_v → Q8_0 (retrieval-critical)
• attn_output → Q5_K
• ffn_gate_inp (router) → Q8_0 (routing-critical)
• ffn_gate_exps / ffn_up_exps / ffn_down_exps (256 routed experts) → IQ4_NL
• ffn_gate_shexp / ffn_up_shexp / ffn_down_shexp (shared expert) → Q6_K
• token_embd / output → Q8_0
• everything else → Q4_K_M (fallback)

Usage

llama-bench -m Qwen3.6-35B-A3B-DYNAMIC.gguf -ngl 99 -fa 1 -ctk q8_0 -ctv q8_0 \
  -ub 2048 -b 2048 -p 4096 -n 128

Benchmark context

Research series on pushing Qwen3.5/3.6 on AMD Strix Halo. Methodology, scripts, and live results: see the benchmark site referenced from the GitHub repo.

License

Apache 2.0 (inherited from base model).