96GB+48GB VRAM all-local agent harness owner compares long-horizon models, asks for picks
CharlesStross · reddit · 2026-10-17
Reddit user CharlesStross shares a fully local agentic setup: a self-built harness with tools, solid web browsing, a skills framework, and agent-managed crons, running on RTX Pro 6000 (96GB) + 2×3090 (48GB) + 128GB RAM.
Real uses: tracking refunds, monitoring easy-to-miss emails (trial renewals), mining medical records before appointments, building Instacart orders, trialing Amazon/eBay purchases, and experimenting with real estate/government auction tracking; two Muse agents serve as free delegated compute.
Models: main driver Qwen Flash Next q6, roughly tied with DSv4F+vision q8 for agentic work; Qwen 3.6 35ba3b + Ninfer hits 500 tok/s for dumb-but-fast tasks. Pain points: long autonomous runs hit 250+ tool calls per turn; Qwen 3.8 27b struggles as tasks lengthen; MiMo codes well but falters when driving a browser, updating memories, scheduling, and emailing in one turn; lighter quants get careless about directives from 80k tokens ago. Author asks fellow all-local users for model picks and long-horizon hacks.
More from coding & agent
- Redditor Questions Muse, the #1 Downloaded App: Are AI Agents Ordering Uber Eats a Real Use Case? — eat-Razzmatazz1809 · 2026-10-17
- Ilya Razensteyn interviewed ~2 dozen engineers on coding agents; writeup due next week — thomasahle · 2026-10-17
- AIPOCH open-sources Open-Science: desktop app where AI agents run reproducible scientific analyses — rohanpaul_ai · 2026-10-17
- Stripe launches real-time Postgres sync for Stripe data, no webhooks needed — alexmacgregor__ · 2026-10-17
- MiMo-V2 paper: agentic groupwise grading redistributes RL advantages to curb reward hacks — Sauers_ · 2026-10-17
- Creator shares full AI music video pipeline: Claude for lyrics, Suno for song, local LTX for video — elonmusk · 2026-10-17