My TensorFold patch makes Qwen3.8-27B process prompts about 10% faster on an M5 Ultra,…
This is a dev post classified by AGK Intelligence as AI dev tools (a tool drop), kept by the Dev Radar because it carries real work, not commentary.
My TensorFold patch makes Qwen3.8-27B process prompts about 10% faster on an M5 Ultra, with identical tokens. Mac Studio M5 Ultra (256 GB), TensorFold before my patch: Prefill, 2K prompt: 1,722 tok/s Prefill, 8K prompt: 1,679 tok/s Prefill, 32K prompt: 1,517 tok/s Prefill, 64K prompt: 1,173 tok/s Mac Studio M5 Ultra (256 GB), TensorFold with my patch: Prefill, 2K prompt: 1,903 tok/s Prefill, 8K prompt: 1,855 tok/s Prefill, 32K prompt: 1,669 tok/s Prefill, 64K prompt: 1,313 tok/s Model: Qwen3.8-27B, MLX 4-bit, about 4.5 bpw (our calculation) DFlash2 drafter: not used in my runs Builds: main
Posted by Niklas (2.3k followers) 1 h ago · 20 likes · 858 views · view the original post on X. Kept by the Dev Radar as AI dev tools. Tools mentioned: tensorfold.
More dev work like this
- Stop copying tasks between your AI agents’ chat windows. — @DanKornas
- The speed of @deepseek_ai V4.1 Flash running of four DGX Sparks is so addictive 🤩 — @javierbits
- THIS IS MASSIVE. THE ONE EVERYONE'S BEEN WAITING FOR 📱🔥 — @HermesWatcher
- i've been a huge fan of @claudeai Code Projects, and now they are making it clear you… — @theaaron
- This is incredible. — @ashxhart
- That reminds me, I forgot to send yall my update of Mica!! v2.5 — @markfenner
- Sails AI can now listen and speak 🎙️ — @Dominus_Kelvin
- Claude Sonnet 5.5 vs Gemini 4 Argon for computer use. — @coastyai
Every post is read and classified by AGK Intelligence: what it is, which market it belongs to, and whether the link is a real tool. 31.2k posts from 5k X accounts over the last 21 days, 8.9k tools, 12 markets. New posts are collected several times an hour and the full ranking is recomputed every hour — last update 2026-10-31 20:27 UTC. Full method.