Tencent's Hunyuan-A13B technical report says the model was pretrained on a 20-trillion-token corpus
26 giu 2025 · Raccolto automaticamente da fonti pubbliche. Non revisionato da una persona.
Questo contenuto è disponibile solo nella lingua originale (inglese).
What happened
Tencent's Hunyuan team states in its Hunyuan-A13B Technical Report, dated 26 June 2025, that Hunyuan-A13B is an open-source mixture-of-experts language model with 80 billion total parameters and 13 billion activated parameters per input, pretrained on a rigorously filtered 20-trillion-token corpus curated from diverse domains, including 250 billion tokens of high-quality STEM pre-training data.
Limit of the disclosure
The report describes the corpus by scale, filtering and curation pipeline (deduplication, low-quality filtering, denoising, topic labeling, and the reused Hunyuan-TurboS pipeline) but does not disclose the corpus's constituent works, rightsholders, or licenses. This item records only the officially disclosed training scale; it does not claim any specific work was included.
Fonti
- Hunyuan-A13B Technical Report — Tencent Hunyuan Team (primary)