Perplexity Flash matches Wald's accuracy at nearly 2x speed in local agent-decision benchmark
WolframRvnwlf · x · 2026-10-17
Testing Perplexity's new local model pplx-decider-flash-v1.2 on a 16 GB M4 Mac mini, the author found it matched Wald — the previous winner of a 14-model benchmark — on accuracy while nearly halving latency. In the real-email task, both got all 29 processable cases right, with Flash at a median 1.35s per completion check vs 2.62s for Wald; synthetic benchmarks showed the same scores with 1.8x speed. The piece explains that agent micro-decisions — picking a tool, routing a message, verifying a promised action happened — only need selecting among defined answers and returning probabilities, no generated explanation, so a fast local model suffices. Flash now takes the crown for local email completion.
More from coding & agent
- Engineer: I'm not smarter than AI anymore — time to learn to orchestrate it — KevinNaughtonJr · 2026-10-17
- Open-source OODA pentest agent drives Kali and Playwright tools directly over MCP — irraz_rulez · 2026-10-17
- DeepSeek V4.1 Flash 1017 finds 54 vulns for $7.53, topping open-weight leaderboard — zizhpan · 2026-10-17
- Claude Code wrote 34,000 lines of a Godot tycoon game, Blender models included — flockbaygames · 2026-10-17
- Maker wires customer testimonial library to Claude over MCP with voice-note capture — Puzzleheaded_Dog3391 · 2026-10-17
- Peter Yang: Software isn't dead — agents are just becoming its main users — petergyang · 2026-10-17