DeepSeek V4.1 Flash 1017 hits 1,000 tok/s on Vercel while a home 4x 3090 rig does 65
AccBalanced · x · 2026-11-09
Vercel's AI Gateway now serves deepseek-v4.1-flash-1017-ultrafast at over 1,000 tokens per second for coding, reasoning, and agents.
Reposting, superalesha contrasts this with the same open model running at 64-68 tok/s decode on his home 4x 3090 rig with community engines and his own patches. In weeks the model went from a Hugging Face upload to cloud 1,000 tok/s, zero-retention hosting, and an uncensored variant on US servers — all without DeepSeek's permission. He endorses Yuchen Jin's take that DeepSeek is an S-tier lab for designing architecture around inference efficiency and open-sourcing all of it: 65 tok/s beats reading speed; 1,000 tok/s is for agents.
Related event: DeepSeek V4.1 Flash hits Vercel AI Gateway claiming 1000 tokens/sec(2 posts)→
More from Infra
- Day 309/365 of GPU Programming: Optimizing LLaMA2-7B on Edge with SIMD and Multithreading — NandoDF · 2026-11-09
- Waymo explains at Hot Chips why camera-only autonomy plateaus at 99.9% — 4310sy · 2026-11-09
- Running a Home Mac mini as an Always-On Node for Remote AI Tasks — xiaohu · 2026-11-09
- GitHub investigates delayed Actions runs and webhook deliveries in new incident — DanielLockyer · 2026-11-09
- Hardware-aware LDM implementation cuts memory traffic via state unwinding and windowed writes — SonglinYang4 · 2026-11-09
- Exact Relative-Position Attention Kernels Deliver Up to 21x Backward Speedup at 8K Context — INSAIT-Institute · 2026-11-09