First hands-on: Meta's Muse Wonder 1.0 vs DeepSeek V4.1 Flash in Codex CLI
Richidm · reddit · 2026-11-03
The first public hands-on of Meta's Muse Wonder 1.0 vs DeepSeek V4.1 Flash (via Codex CLI): same production TypeScript task (19 files, 330 lines), same repo and prompt, snapshot-restored between runs, with Claude reviewing code line by line instead of trusting self-reports.
Results
- Time: Muse 47 min wall (with one human-needed stall, 24 min working); DeepSeek 29 min, zero interventions
- Tests: Muse 1728/1728 with 10 new tests; DeepSeek 1726/1726 with 8 new, plus a proper 'red' run of 7 behavior-missing failures first
- Mutation testing (40 mutants): 7 vs 6 killed; both produced correct, visually identical video renders
Code quality
- Muse: two real bugs tests missed — copy-pasted a 50-line background-removal function so every video runs it twice (doubling rented-GPU cost); ignored the job-id-seeded random fallback spec, seeding on title instead; hard-coded the closed name list in two drifting places
- DeepSeek: cleaner structure (refactored into a single loop, no duplication) with three small slips
Verdict: on-screen output equally good, but DeepSeek's code is notably better engineered while Muse is faster but leaves hidden costs and maintenance traps.
More from coding & agent
- TesterArmy rebrands as e2e, raises $8M seed for agentic coding verification layer — ycombinator · 2026-11-04
- Grok bills bots by steps, not messages — 8 tips to stop burning your usage — nima_owji · 2026-11-04
- Code verification startup raises $8M, hits 1.5M npm downloads in first month — lukas_m_ziegler · 2026-11-04
- Open-source openwiki ships claims, agent integrations and cross-repo wiki linking — BraceSproul · 2026-11-04
- Merge launches Universal Context Layer: one MCP-connected context brain for all agents — shensi · 2026-11-04
- LocalVocal: fully offline Mac app lets you voice-chat with local LLMs and agents — joshwhiton · 2026-11-04