Michał Piszczek

cdiamond

AI & ML interests

CTO of Archdesk · Founder: Lextron.ai, Inclify, Rejsomat.pl, Robotero · Autonomous AI agents · AI reliability · Software architecture · ConstructionTech · Proof-Adjusted Autonomy · Joule Wars

Recent Activity

repliedto Avifenesh's post 3 days ago

Wrote this up because it still feels backwards. Acceptance dropped, tok/s went up. Full draft head: 66.7% / 117.1 tok/s. Trimmed: 63.6% / 121.7. Later remeasure +5.1% on a Pro 6000, +6.4% on a 5090 laptop. Full head still accepts more. Loses anyway. Not my idea. FR-Spec. I just wired it for the native MTP head. https://huggingface.co/blog/Avifenesh/masked-mtp-drafts https://huggingface.co/Avifenesh/Qwen3.8-27B-NVFP4-MTP-GGUF

new activity 3 days ago

cdiamond/Qwen3.8-27B-iMatrix-NVFP4-MTP-GGUF:Why the NVFP4 MTP head beats a higher precision one (measured)

updated a collection 3 days ago

Qwen3.8 27B at 256K on 24 GB

View all activity

Organizations

New activity in cdiamond/Qwen3.8-27B-iMatrix-NVFP4-MTP-GGUF 3 days ago

Why the NVFP4 MTP head beats a higher precision one (measured)