Jackrong's Qwopus3.6 Hits 161 Tokens per Second on a Single Consumer GPU
A community fine-tune of Alibaba's Qwen3.6 MoE hits 161.9 tokens per second on a single RTX 5090, with a 262k context window and one-shot frontend generation.
Takeaways
- Qwopus3.6-35B-A3B-v1 ships in GGUF, fine-tuning Alibaba's Qwen3.6 MoE with three-stage SFT.
- 35B total parameters, only 3B active per token via 256 experts and Gated DeltaNet attention.
- Hits 161.9 tokens per second on a single RTX 5090, a 2.6x speedup over the 27B dense predecessor.
- Native 262k context; long-context stability improves noticeably with YaRN RoPE scaling.
- Scored 88.6 overall, 94.2 quality, 91.7 reliability in independent GGUF benchmarks.
- Apache 2.0, but weight merging with PEFT and Unsloth is buggy; treat as experimental.
A new community fine-tune landed on Hugging Face that turns Alibaba's Qwen3.6-35B-A3B base into a reasoning-tuned local model with unusually strong one-shot code generation. Called Qwopus3.6-35B-A3B-v1 and shipped in GGUF format by Jackrong, the release has already crossed 135k downloads and feels like a much larger dense model while running on a single consumer GPU.
The base architecture is what makes this possible. Qwen3.6-35B-A3B is a hybrid sparse Mixture-of-Experts model with 35B total parameters but only 3B active per token, routing across 256 experts and combining Gated DeltaNet linear attention with standard gated attention layers, with a native 262k context window. Gated DeltaNet is a linear-attention variant whose memory scales almost flat with sequence length, so the model can hold roughly 200k words in context without VRAM exploding.
Throughput on a single 5090
Independent evaluation shows Qwopus3.6-35B-A3B-v1 leading in overall quality and reliability while maintaining strong inference speed on consumer hardware. On an RTX 5090 running the Q5_K_M quant, it clocks about 161.9 tokens per second, a 2.6x jump over the previous 27B dense Qwopus preview. In a comparison table against similarly sized fine-tunes, it scored 88.6 overall, 94.2 on quality, and 91.7 on reliability.