Dev Runs DeepSeek V4 Flash Locally at ~1400 tok/s, Fixes Two iOS Bugs End-to-End
HankYeomans · x · 2026-09-02
Developer HankYeomans tested a locally deployed DeepSeek V4 Flash (MXFP4, claimed 1M context) as a coding agent.
- Found the Claude harness caps 'unknown models' at 200K context, so he ran it via LM Studio instead.
- After joint planning, the model autonomously executed phases and handled commits end-to-end; with a few skills it fixed two iOS bugs without CI, linting, or Xcode testing set up.
- It nearly maxed out two GPUs' VRAM; LM Studio logs showed 1400 tokens/sec, and he had Amp (Ultra) review the work.
- He says the experience rivals pre-release frontier coding agents. Note: DeepSeek V4 Flash is not yet officially confirmed.
Related event: Dev Runs DeepSeek V4 Flash Locally on Dual GPUs to Fix iOS Bugs(2 posts)→
More from coding & agent
- Claude-written test harness let 31KB fake archives land on a real backup server — Grimmoner · 2026-10-18
- Open-source Tyger-S2S voice agent runs natural, interruptible conversations on a single GPU — solyarisoftware · 2026-10-18
- Matt Pocock's pre-commit hook caps agent-written files at 5,000 tokens — mattpocockuk · 2026-10-18
- Backups silently failed for 10 days: three monitoring blind spots lined up — Grimmoner · 2026-10-18
- Author Laurence Moroney: write your ugly first draft yourself, then let an LLM edit it — lmoroney · 2026-10-18
- Open-weight TYGER speech decision model makes 80ms turn-taking calls for full-duplex voice agents — solyarisoftware · 2026-10-18