Browser-native Coding Agent: MiniCPM5-2B 4-bit ONNX runs agent stack fully in-browser at 25 tps
karminski3 · x · 2026-09-15
A Hugging Face Space project, MiniCPM5-2B-WebGPU-Pi, packs a full coding agent into the browser: the Pi framework wraps the agent, driven by a sub-2GB 4-bit ONNX build of MiniCPM5-2B, so the entire stack — framework and inference — runs locally in the browser. No Docker, Node, or API keys needed; opening the page gives you a virtual terminal with file editing and tool calls.
The author benchmarked 25 tps on an RTX 3080 Ti and found the 2B model's tool calling surprisingly stable. Two novel ideas emerged:
- Inherited login state: since the model runs in your browser, the agent naturally carries your cookies and sessions — useful for automating internal-only SaaS or private network projects, even offline.
- Self-healing web pages: listen for console errors and hot-patch the page in place, no refresh required.
The author is considering a side-model agent benchmark roundup if there's interest.
Related event: MiniCPM5-2B Runs a Full Coding Agent in the Browser via WebGPU(3 posts)→
More from coding & agent
- Wiring Unreal, Comfy and Blender MCP Servers Into One Coding Agent to Build a Game — PurzBeats · 2026-09-15
- Dev + SWE-2 High Fusion Mode Inside Devin Called a 'Holy Grail' Combo — CtrlAltDwayne · 2026-09-15
- Operators say agent-accelerated teams still need human judgment for strategy — infoxiao · 2026-09-15
- 12.9k-star Claude skill writes accurate prompts for any AI tool, zero wasted retries — tom_doerr · 2026-09-15
- OSS harness boosts OS World 2.0 score ~20%, jumping Sol/Max from 4th to 1st — demeyer1 · 2026-09-15
- What should NOT be an MCP server? Developers debate where MCP is overkill — JuicerSocial · 2026-09-15