I built an open-source agent harness that lets a 2B model finish real tasks (288-cell benchmark, all data public)

Description and use

Mingbird is an open‑source agent harness that lets a 2‑B model complete real tasks, achieving 82 % success on 18 tasks, comparable to a 35‑B model.

Problem

Small 2‑B models often fail to complete real tasks reliably, looping on tool calls or declaring victory prematurely.

Outcome

Mingbird enabled a 2‑B model to complete real tasks with high reliability, closing the performance gap with larger models and outperforming other harnesses on a 288‑cell benchmark.

Workflow

• Run 18 real tasks with a 2‑B model using Mingbird harness
• Measure success via deterministic scoring of produced artifacts
• Compare results against other harnesses and larger models

Evidence

Every person who has tried to make a 2B model do real work on their own machine knows the failure mode. The model loops on the same tool call, or declares victory with half the deliverables missing, or just goes silent in the middle of the task.

Source

Mingbird handles that layer, and a 2B model finishes real work.

Source

Four agent harnesses (Mingbird, goose, opencode, agent-mini, all stock, unpatched) × four open models (gemma4:e2b 2B, qwen3.5:4b, gemma4:12b, ornith-1.5:35b) × 18 real tasks = 288 cells.

Source

Overall: 0.886 / 0.631 / 0.479 / 0.405.

Source

At 2B, three of the four harnesses fall off a cliff and Mingbird does not: 0.821 is within 0.12 of its own 35B score.

Source

One task (WF-08), same 4B model, two builds of my own harness: Before one fix: 156 tool calls, zero files produced. After: 15 tool calls, three files.

Source