Ministral 3 3B on a Galaxy S21 relays chats between Gemini and Z.ai across two browsers, 10/10 runs
Mean-Standard7390 · reddit · 2026-09-16
A striking edge-agent demo: Ministral 3 3B (Q4KM) running in llama.cpp on a Samsung Galaxy S21—4 CPU threads, no GPU, on battery—relayed conversations between Gemini in Chrome and Z.ai in Firefox, with no integration between the services. Ten consecutive runs all succeeded: 4 browser hops each, 40 replies, 20 automatic chat resets, median 9:29 per run, 16% battery over 98 minutes.
The key insight is what the model sees, not how smart it is: no screenshots or raw HTML, just a short list of named elements (118 Gemini nodes collapsed to 2 candidates; 124 on Z.ai to 4). A control feeding raw HTML to Qwen3-0.6B used 25x prompt tokens and 17x the time, and a live Wikipedia page didn't fit the phone's context at all. Notable details: the model picked the real send button among decoys 20/20 times, first decisions take 9s vs 3s after, JSON output drifts, and one run needed a recovery retry. Script, hashed JSONL logs, and redaction procedure are on GitHub (e2llm/edge-browser-agent); the author notes it's a single-scenario harness, not a benchmark.
More from coding & agent
- Claude Code bypass permissions mode swaps built-in tools for Bash — stablequan · 2026-09-16
- Dev wishlist: coding harnesses should treat automations as one thing, not 100 chats — blixt · 2026-09-16
- Qwen 27B on a single RTX 5090 generates 8-scene animation from one prompt — Acceptable-Object390 · 2026-09-16
- 37-year-old non-coder builds DnD matchmaking site Tavernly with Claude Code — DopeDutchLemon · 2026-09-16
- ChatGPT's built-in Sites feature turns a single prompt into a live, publishable website — TawohAwa · 2026-09-16
- Blogger builds satellite-imaging financial due diligence tool on Doubao Seed 2.1 — karminski3 · 2026-09-16