Qwen-MT-Lumen
Qwen-MT-Lumen is a released general LLM work model with open-source; evaluate it while provider pricing coverage matures.
- Family
- Qwen-MT
- Released
- 2026-08-07
- Context
- 4K
- Parameters
- 14.8B
- Architecture
- Decoder Only
- Specialization
- translation
- Openness
- Open source
- License
- Apache 2.0OSI-approvedPermitted
- Weights
- Available
- Code
- Unknown
- Training
- Multi-stage
No tracked provider pricing is available yet.
No tracked provider pricing is available yet.
About
Lumen is a 14.8B-parameter dense translation model submitted by Alibaba Token Hub to the WMT 2026 General Machine Translation shared task (constrained track). The Hugging Face card describes it as built on the Qwen3-14B architecture; the system paper (2026.wmt-1.81) describes training from Qwen3-14B-Base with continued pre-training, SFT, preference optimization and on-policy distillation, plus a structure-aware inference pipeline. Weights are public and ungated under Apache 2.0 at ATH-MaaS/Qwen-MT-Lumen (the paper's ATH-MaaS/Lumen_wmt26 URL redirects to that repo). Context is a hard 4096-token limit. It is not Alibaba's commercial Qwen-MT-Flash/Plus/Lite product line.
Providers
No active provider routes are tracked for this model yet.
Capabilities
No model capability flags are currently sourced.
Benchmark peer barsfor Coding
No task-mapped benchmark peers are available for this model yet.
Benchmark scores(1)
| Benchmark | Score | Version | Evaluation | Source |
|---|---|---|---|---|
| 6.8 | WMT26 General MT findings, Figure 3 Avg. Rank (lower is better)Observed 2026-11-10 | Configuration: Lumen (constrained-track system; 14.8B backbone plus the structure-aware inference pipeline in the HF repo) Harness: Contrastive Error Span Annotation (cESA), human annotators Evaluator: WMT26 organizers Confidence: confirmed Notes: WMT26 General MT findings (2026.wmt-1.48), Figure 3: Avg. Rank 6.8 for the system listed as 'Lumen', averaged over the 23 language pairs it entered; rank 1 is best, so lower is better (not a 0-100 score). Official wording: Lumen 'led the constrained track, ranking in the top cluster for 20 of 23 language pairs'; 'Among the constrained systems in Section 6.3, Lumen emerges as the clear winner'; the best system overall was Wayfinder (17 of 23 top-cluster pairs), then Gemini 3.1 Pro (15), TRIVE (14), and Lumen (13). Per-pair cESA scores (0-100, higher is better, not comparable across pairs): en-ja 90.6, en-arz 83.1, cs-vi 85.9, cs-uk 86.6, en-id 82.6, cs-de 80.2, en-uk 81.5, en-zh-Hans 78.6, en-ru 79.8, zh-ja 73.7, en-ko 82.5, en-th 84.0, en-cs 65.5, en-de 72.0, en-zh-Hant 74.0, en-kk 82.2, en-et 80.3, en-is 84.5, en-hy 83.4, en-be 75.1, en-se 54.5, en-lij 25.4, en-lld 34.8. The official score is for the submitted system (backbone + structure-aware inference / retry pipeline described in 2026.wmt-1.81). Source: WMT26 organizers. Retrieved 2026-11-10. | Source Source: Findings of the WMT26 General Machine Translation Shared Task (Kocmi et al., 2026) |
Migration checks
No linked migration route is available for this model yet.
API versions
qwen-mt-lumen