Mellum2.1-12B-A2.5B hands-on: strong agentic coding on a laptop, weak one-shot generation
SoAp9035 · reddit · 2026-10-09
A redditor tested JetBrains' Mellum2.1-12B-A2.5B (Q8) locally with the Pi coding agent and llama-server, all one-shot. SVG/OS/Minecraft generations were poor, but agentic behavior shone: reliable tool calling, self-correction of editing mistakes, successful codebase edits on an existing project, and even converting screen recordings to GIFs autonomously. Runs 40 t/s at 131K context on an 8GB VRAM laptop. Verdict: not impressive at one-shot creation, but genuinely usable for small dev tasks.
More from coding & agent
- Decisions API hits GA as agents automate cloud video editing via feedback loops — dkundel · 2026-10-23
- Swarms API changelog: 8 months of streaming, multi-agent orchestration, and dynamic model pricing — KyeGomezB · 2026-10-23
- A curated library of 50+ free AI agent resources: videos, repos, one Google Doc — ifioknkem · 2026-10-23
- Vercel adds no-account deploys: drag and drop a file to get a live URL — steren · 2026-10-23
- Dev runs 5 Claude Code sessions that coordinate via MCP Agent Mail to root his network cameras — doodlestein · 2026-10-23
- Vercel ships 'vercel deploy --temporary' for ephemeral agent-built deployments — cramforce · 2026-10-23