3 used V100s hit ~400 tokens/s with custom sFNX engine, author defends claim
jfinkpottery · reddit · 2026-10-27
After being accused of lying about near-400 tokens/s speeds, the author explains that MTP gives wildly different speeds per prompt: 376 tokens/s on a realistic "edit this Python file" task (max seen 404), and 140 tokens/s on prose reasoning—still double any other engine on this hardware. Setup: his own sFNX engine (strata/llama-based, engine code rewritten) running flash-next Q4XL on 3 used V100 32GB cards ($600 each) in a decade-old Dell R740 ($500 barebones). Most experts need to fit in VRAM for best results.
More from coding & agent
- Alter Zero: open-source Rust terminal agent in a 24MB binary for coding and security — Hacking0912 · 2026-10-27
- Dev adds a guard before cron jobs: check windows, budgets, leases first — petrusenko_max · 2026-10-27
- golang/go repo flooded with a dozen AI-generated CVE security issues in minutes — evilsocket · 2026-10-27
- SuperPlane 1.0 launches open source: 52% of team's merged PRs were automated — Arindam_1729 · 2026-10-27
- OpenAI DevDay Berlin: Codex, Agents API and Terra 6.1 launch on site — paw_lean · 2026-10-27
- MiniMax M3.1 Flash hands-on: agent builds, tests and iterates a game end-to-end — Aiden_Tech_Ai · 2026-10-27