Mistral Large 4 (1T) on one Mac Studio M5 Ultra decodes code at 22.1 tok/s and needs…
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.
Mistral Large 4 (1T) on one Mac Studio M5 Ultra decodes code at 22.1 tok/s and needs 242.3 s to the first token at 32K Mac Studio M5 Ultra (256 GB): Decode, code: 22.1 tok/s Decode, prose: 25.2 tok/s First token, 2K: 9.1 s First token, 32K: 242.3 s 1 at once: 18.4 tok/s 4 at once: 25.1 tok/s 8 at once: 25.3 tok/s The model cannot turn thinking off, so every answer cut at 512 tokens ends inside its reasoning, and decode counts those tokens. 8 at once is above the server's 4 slots. Unsloth UD-IQ1_S, about 1.8 bpw (our calculation), llama.cpp b12020 with the DSpark draft, context 65,536. Main
Posted by Niklas (2.4k followers) 1 h ago · 3 likes · 225 views · view the original post on X. Kept by the Dev Radar as AI dev tools.
More dev work like this
- Mostly agree with @andyyy here. — @rgvrmdya
- Training a Simple Neural Network! BigData #Analytics #DataScience #AI #MachineLearning… — @gp_pulipaka
- #ReinforcementLearning with #TensorFlow! #BigData #Analytics #DataScience #AI… — @gp_pulipaka
- Switch interfaces without restarting your agent conversation. — @DanKornas
- Claude Sonnet 5.5 vs GPT-6.1 Sol for phone use. — @coastyai
- DeepSeek-V4.1-Flash-1017-GSQ-RCO-GGUF is here! This quantized MoE model packs serious… — @HuggingModels
- Run agent code on your iPad — not just chat about it. — @DanKornas
- DeepSeek just open-sourced its Super-agent prototype. One router session talks to people… — @kimmonismus
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 2026-11-01 17:07 UTC. Full method.