Qwen3.6 35B (3B active)
on NVIDIA RTX 3060 12GB · Strata
Use cases
codingcreative-writing
Summary
User reports Qwen3.6-35B-A3B at 118 tok/s decode on an RTX 3060 12GB with the Strata engine, against 68 tok/s for llama.cpp.
Setup is Strata with Q2_K_XL weights at 11.7 GiB, so some experts stay on the CPU; the machine has a Ryzen 5 7600X. Prefill is about 2200 tok/s at 4K, and a 32K context gives 1800 prefill and 92 decode.
With MTP decode reaches up to 160 tok/s (152 after a 4K prompt), which the user calls a bonus on top of the plain numbers; MTP gave no gain in llama.cpp on this card. The user notes MTP gains depend on the text, with code benefiting more than creative writing, and that Vulkan matches CUDA speed.
Compare this run
Medians of plain runs: one device, one request, no speculative decoding, the whole model in the device's memory. The report above is one setup, so read it beside them and not as one of them.
Qwen3.6 35B · 3B active at 4-bit by hardware
- NVIDIA RTX 509032 GB215.1 t/s · one run
- 178.0 t/s · one run
- 137.7 t/s · 2 runs
- AMD RX 7900 XTX24 GB125.5 t/s · one run
- NVIDIA RTX 309024 GB101.7 t/s · one run
- 81.1 t/s · one run
Bars are medians on one scale, 0 to 215.1 tokens/s. Each dot is one run and opens its report. Press ▶ to watch text arrive at a row's speed.
Models most run on the NVIDIA RTX 3060 12GB
- Gemma 4 8B4-bit59.5 t/s · 2 runs
- Gemma 4 5.1B4-bit60.0 t/s · one run
- Qwen2.5 7B4-bit49.0 t/s · one run
- 113.6 t/s · one run
Bars are medians on one scale, 0 to 113.6 tokens/s. Each dot is one run and opens its report. Press ▶ to watch text arrive at a row's speed.
Everything reported on the NVIDIA RTX 3060 12GB →Every Qwen3.6 report →Estimate Qwen3.6 on this card →