DeepSeek-V4.1-Flash-1017 NVFP4: 4 DGX Sparks decode code at 33.6 tok/s and reach 114.8…
This is a dev post classified by AGK Intelligence as Hosting & infra (a tool drop), kept by the Dev Radar because it carries real work, not commentary.
DeepSeek-V4.1-Flash-1017 NVFP4: 4 DGX Sparks decode code at 33.6 tok/s and reach 114.8 tok/s with 8 requests at once 4 Sparks (512 GB), NVFP4: Decode, code: 33.6 tok/s Decode, prose: 30.2 tok/s 500 words: about 22.4 s First token, 32K: 8.5 s 1 at once: 27.6 tok/s 4 at once: 73.5 tok/s 8 at once: 114.8 tok/s On these prompts the draft gives 2.5 tokens per step. On one long HTML answer the same server gave 4.95 and decoded 63.3 tok/s. NVIDIA's checkpoint with ondigo-winder's recipe on vLLM, tensor parallel 4, about 4.6 bpw (our calculation), context 262,144. Main changes from the recipe: NVI
Posted by Niklas (2.5k followers) 2 h ago · 6 likes · 584 views · view the original post on X. Kept by the Dev Radar as Hosting & infra.
More dev work like this
- I plugged a 100-gigabit datacenter network card into a MacBook Pro running Linux. — @JoshuaSWarren
- that's it — @AniC_dev
- HLD Interview Notes: Day 13/30 — @SCR01111
- P99 CONF 2026 | From Snakes to Crabs: Rewriting Feature Flag Evaluation Infrastructure…… — @rustaceans_rs
- Run GLM 5.3 Flash on your @NVIDIAAI 2x RTX PRO 6000s with ease, on TensorFold 🔥 — @MiaAI_lab
- 𝐀𝐳𝐮𝐫𝐞 𝐂𝐥𝐨𝐮𝐝 𝐌𝐢𝐧𝐝𝐦𝐚𝐩 - A Practical Guide for Data & Cloud Engineers ☁️ — @AiswaryaVenkit1
- Feature flags flat out changed how I ship. — @theayush
- Need help in an unattended store? Raise your hand.🙋 — @seeedstudio
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. 31.3k 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 20:41 UTC. Full method.