Cola2
ByteDance Seed🇨🇳 CNMajor lab· Nov 6, 2026
Textopen weights
Cola2 is a base text generation model that compresses text into semantic spans, generates continuous latents with a diffusion transformer, and decodes them into tokens. The model card reports a 60.10 average over eight tasks after training on 346.03B source tokens.
Summary of the model card
- Modalities
- text → text
1 variant
| Standard | HF |
Sources
HuggingFacePaperDeutsch
Cola2 ist ein Basismodell für Textgenerierung, das Text in semantische Spans komprimiert, kontinuierliche Latents per Diffusion Transformer erzeugt und diese in Tokens dekodiert. Laut Modelcard erreicht es im Schnitt 60.10 Prozent über acht Aufgaben bei 346.03B Quelltokens.
Benchmarks
| Benchmark | Domain | Value | Evidence |
|---|---|---|---|
| TruthfulQAvendor number | safety | 63.04 % wahrheitsgemaess | 48 |
| MMLUvendor number | knowledge | 63.19 % correct | 39 |
Vendor-reported numbers are marked and count with factor 0.7 in rankings. The evidence score says how much a benchmark still tells you today.
Provenance
- Detected
- Nov 9, 2026
- First source
- huggingface
- Description
- Summary of the model card
- Editorially reviewed
- —