alt.hn

4/22/2026 at 1:36:43 PM

Qwen/Qwen3.6-27B · Hugging Face

 https://huggingface.co/Qwen/Qwen3.6-27B

by cgeier

4/22/2026 at 2:38:04 PM

Thanks to unsloth, we immediately get 4-bit quantization: https://huggingface.co/unsloth/Qwen3.6-27B-GGUF

by throwaw12

4/22/2026 at 3:04:59 PM

Im waiting for nvfp4 quant :-)

by perarneng

4/22/2026 at 8:14:18 PM

Are nvfp4 / mxfp4 even useful without QAT?

by EnPissant

4/22/2026 at 1:45:33 PM

According to (their own) benchmark, this is beating Opus 4.5. Hard to believe for such a small model.

by cgeier

4/22/2026 at 2:26:59 PM

This is an amazing win for the open source but the terminal bench 2.0 (only benchmark I am familiar with) scores are 'non standard'

> * Terminal-Bench 2.0: Harbor/Terminus-2 harness; 3h timeout, 32 CPU/48 GB RAM; temp=1.0, top_p=0.95, top_k=20, max_tokens=80K, 256K ctx; avg of 5 runs.

Terminal bench 2.0 rules explicitly disallow modifying timeouts or resources available. Each terminal bench task has timeout (usually under 1h, mostly under 30 mins) and resources configured in the docker container by the task creator and they are chosen that way to test specific model aspects.

by GodelNumbering