Marathoner: a 9B model that codes for 10+ hours, lifting SWE-bench Verified to 77.5%
mark_k · x · 2026-10-01
Marathoner: persistent coding in a 9B model
- Researchers from Ant Group, Peking University and University of Macau report Marathoner, built on Qwen3.5-9B, capable of 10+ hours of continuous coding and 1,000+ tool calls.
- Training: software tasks mined from 100K GitHub PRs across 10K repos, chained for difficulty; trained on successful Kimi K3 runs; RL in real execution environments; a bonus specifically rewards useful late-run progress like finding hidden bugs.
- Results: SWE-bench Verified 43.8% → 77.5%; Terminal-Bench 2.0 27.3% → 57.2%. Still behind frontier models, but a huge gain from the same 9B start.
- Author's take: an agent that keeps making progress while you sleep changes what you can delegate — and this persistence can be trained into small models.
Related event: Ant Group Unveils Marathoner, an Agent That Codes for 10+ Hours Straight(2 posts)→
More from coding & agent
- Generic API-catalog MCP server vs one tool per job: a context-bloat experiment — MandeepKaur9407 · 2026-11-16
- Berkeley paper: LLMs answer with stale values even when the update is in context — ombre-ai · 2026-11-16
- skills-gate: one gate skill fronts 148 agent skills at just 389 tokens — Distinct-Pie2389 · 2026-11-16
- Open-weight Karpo-Decision-31B tops Decision Index at 73.9, one forward pass per answer — CodeByPoonam · 2026-11-16
- shadcn boosts @uisdk: one-shot UI rebuild that's instantly E2E-testable by AI — shadcn · 2026-11-16
- AI agent deletes 43,599 lines of OpenClaw test code in one PR — vincent_koc · 2026-11-16