Mtp on Marco Scavazzonhttps://scavazzon.com/tags/mtp/Recent content in Mtp on Marco ScavazzonHugoen-gbWed, 22 Jul 2026 11:36:11 +0200Qwen 3.6 27B: MTP Benchmark on a 24 GB GPUhttps://scavazzon.com/posts/qwen-3.6-27b-mtp-benchmark-on-a-24-gb-gpu/Wed, 22 Jul 2026 11:36:11 +0200https://scavazzon.com/posts/qwen-3.6-27b-mtp-benchmark-on-a-24-gb-gpu/Qwen 3.6 27B MTP benchmark on a 24 GB RTX 3090. Speculative decoding nearly doubles code-generation speed, with +44% on reasoning and clear VRAM trade-offs.