60 real PRs put Muse Spark Open to the test: 48% bug-fix rate, 72% clean PR approval
Arindam_1729 · x · 2026-10-20
Instead of the usual one-toy-prompt vibe review, the author gave Meta's new open-weight coding model Muse Spark Open (816B total / 42B active params, 1M-token context, billed as the most capable US open-weight model) 60 real PR tasks across 9 repos: django, svelte, tldraw, ant-design, prettier, recharts, tiptap, vuejs/core and mantine.
- Bug PRs: 12 of 25 caught (48%), plus 3 partial. Django was best at 55% (6/11); svelte, recharts and tiptap went 0%.
- Clean PRs: 18 of 25 approved (72%), with tldraw and ant-design at 100%.
The article details where the model shines and struggles on real codebases.
More from coding & agent
- uv author explains: env markers as decision diagrams, minimal equivalent DNF is computationally hard — charliermarsh · 2026-10-20
- uv's scaling pain: complex resolutions explode the intermediate DNF, clause comparison gets expensive — charliermarsh · 2026-10-20
- Yeufonic: open-source frontend for YuE2/ComfyUI aimed at musicians, with local LoRA training — AcrobaticMaize2408 · 2026-10-20
- OpenCode ships DeepSeek V4.1 Flash 1017 update, zero changes needed for users — intellectronica · 2026-10-20
- "The LLM is the brain, the harness is the body": framing agent engineering — EricBuess · 2026-10-20
- Coinbase lets AI agents trade perps and index products via x402 protocol — MurrLincoln · 2026-10-20