Qwen3.6 64K real-world test

Running Qwen3.6-35B-A3B UD Q4_K_S GGUF on an RTX 5070 12GB via a llama.cpp OpenAI-compatible server, the 64K context + thinking setup actually worked.

The model file was 19.45 GiB, with total parameters of 34.66B and active parameters of A3B, and in the MoE structure, 8 experts out of 256 experts were used per token.

Memory usage was roughly at the following levels.

• CUDA model buffer: about 9.46 GiB
• CPU mapped model buffer: about 11.32 GiB
• KV cache 64K: about 465 MiB
• Compute buffer CUDA: about 1.97 GiB

Performance varied significantly by range.

• 10K~20K prompt: prompt eval 1,420~1,480 tok/s, generation 41~47 tok/s
• ~30K context: generation 39~43 tok/s, still usable in practice
• ~40K context and above: generation drops to 12~14 tok/s, and long responses become heavy

Also, the message forcing full prompt re-processing due to lack of cache data appeared repeatedly, and cache reuse sometimes failed to work well when there were large prompt changes.

In summary, 16K~32K is the range that feels best in practice, and while 64K is possible, it's too heavy to use as a default preset for speed purposes.