Qwen-MT-Lumen wins 18 of 23 constrained tracks at WMT26; Wayfinder named Best System Overall
千问大模型 · wechat · 2026-10-15
- At WMT26, Qwen-MT-Lumen (14B) won 18 of 23 constrained-track directions with 3 runner-up finishes, rated "Clear Winner"; Alibaba's Wayfinder took 5 wins and 4 second places and was named the only Best System Overall in the unrestricted track.
- Qwen was the most-used base model family among submissions (9 systems, vs 6 each for GPT and Gemma families).
- WMT26 General MT spans 23 directions and 19 languages with domain instructions and structural constraints, evaluated via cESA human comparison over 46 systems.
- Lumen's four-stage pipeline: multilingual CPT+SFT, online M2PO preference optimization, on-policy distillation from Qwen3-32B-Instruct (teacher used only in training), and structure-aware post-checking — backed by data flywheels from 1B+ daily translation calls.
- Wayfinder uses multi-agent orchestration (candidate generation, quality assessment, anonymous selection, constrained fusion), reaching Top Cluster in 17 of 23 directions.
More from Models
- Beam's Ten-Benchmark Breakdown: Doubles Peers on CritPt, Near-Bottom on Terminal-Bench — ArtificialAnlys · 2026-10-16
- Beam Scores 28 on Coding Agent Index, 40 Behind Leaders but Among Most Token-Frugal — ArtificialAnlys · 2026-10-16
- Beam Token Breakdown: Only 26% of Its 26k Output Tokens per Task Are Answers — ArtificialAnlys · 2026-10-16
- Reflection's First Model Beam Scores 30, Ranks #2 Among US Open Weights Models — ArtificialAnlys · 2026-10-16
- Why are sub-15B open models so rare now? Labs seem to have settled on 24-27B minimum — Hot_Example_4456 · 2026-10-16
- User spots model openly generating "ideal responses" to train other models — liminal_bardo · 2026-10-16