Apex-35B on one 4090: 205 tok/s, self-repairing a Python project in 21 minutes
lxfater · x · 2026-11-03
A deep hands-on test of the new Apex-35B-v1.0: on a 48GB RTX 4090, the author burned 4M tokens across nine benchmarks and concluded it's in a league of its own as a single-card research agent. Key numbers: via Strata it hits 115 tok/s with just 24GB VRAM (72% faster than llama.cpp), 205 tok/s at full 48GB; on a 64-vessel fluid optimization problem it independently derived the flow-radius 3.5-power law within 30 rounds, shrinking error 580x; wired into Claude Code it repaired a broken Python project in 21 minutes with 55 tool calls, passing all 18 tests unattended. lxfater, who previously doubted local small models could do real work, says ordinary single-GPU results are still unverified.
More from coding & agent
- A one-night AI-built project nets 333 points and ~22k visitors — Aizkmusic · 2026-11-03
- Five manual checks before trusting an agent actually finished its job — Kodenstudio · 2026-11-03
- Open-source AI pit crew for iRacing runs 5 agents on Kimi, checks every number against telemetry — JeremyCMorgan · 2026-11-03
- Agent control plane benchmark: Smalltalk-80 cascades hit 100% compliance vs 2.5% for Gherkin — RandalSchwartz · 2026-11-03
- Karpathy: one prompt to Claude Code deploys a YouTube thumbnail-swap browser extension — amaarora · 2026-11-03
- Cherry Pi proposes layered harness architecture: agent loops compose with governed workspaces — sull · 2026-11-03