Tencent Hunyuan-Large

Original Article ↗ Hacker News Discussion ↗

Model architecture & performance

  • Hunyuan-Large is a Mixture-of-Experts (MoE) model: ~389B total parameters, ~52B active per token, 256K context.
  • Benchmarks in the thread say it outperforms Llama 3.1 70B and is comparable to Llama 3.1 405B, despite far fewer active parameters.
  • Commenters note the “significantly larger” language is somewhat misleading since 405B vs 389B is only ~4% difference in total, but MoE means runtime cost is closer to a ~70B dense model.
  • Some see this as evidence of rapid efficiency gains; others question whether this is genuine progress or just more complexity/inefficiency.

MoE mechanics & hardware requirements

  • Architecture: 16 experts with 1 chosen per token plus 1 shared expert always active → ~52B active params.
  • Inference speed depends heavily on batch size; with batch size 1, cost is close to active params, while larger batches can approach full-param cost.
  • To run locally, entire 389B needs to fit in (V)RAM for usable speed. Swapping experts over PCIe would drop to ~1–2 tokens/sec.
  • Rough guidance: ~1 GB of RAM per 2B parameters at 4-bit quantization → ~256 GB RAM plus at least one GPU for practical use.
  • Advanced setups discuss sharding across multiple GPUs/nodes and even large CPU-only rigs with high-bandwidth DDR5, achieving usable speeds on big MoE models.

“Open source” claims & licensing restrictions

  • The project and accompanying paper call the model “open-sourced,” but the license:
    • Excludes the EU entirely from the territory.
    • Imposes an Acceptable Use Policy.
  • Multiple commenters argue this conflicts with the Open Source Definition (discrimination by user and by field of use), so it should not be labeled open source.

EU exclusion & regulatory context

  • License explicitly excludes the European Union.
  • Explanations offered:
    • Avoiding GDPR, the AI Act, and obligations for “systemic risk” models (e.g., disclosure, evaluations, incident reporting, cybersecurity).
    • Possible training on data that would trigger EU privacy issues.
  • Some defend avoiding EU legal exposure; others see EU protections as a feature, even if it limits model access.

Copyright, model weights & ethics

  • Extended debate on whether model weights are copyrightable:
    • Comparisons to phone books, encyclopedias, and EU-style database rights.
    • Distinction between US and EU/UK approaches to collections of facts.
    • Discussion of whether weights are “just facts” vs creative probabilistic structures influenced by many hyperparameter choices.
  • Legal concepts raised: substantial similarity, independent creation, fair use, and potential future “sui generis AI model weights” rights.
  • Ethical concern: even if current law is unclear, many see training on others’ work without compensation as exploitative and argue the law should change.

Broader implications & attitudes toward AI

  • Some see this release as thrilling evidence of rapid capability growth (MoE, distillation, synthetic data, etc.) and celebrate local, powerful models that can meaningfully assist with coding and analysis.
  • Others are skeptical, emphasizing hallucinations, low-quality outputs, and fear of disempowerment and job replacement.
  • Counterarguments highlight personal productivity gains and the view that restricting AI (e.g., via heavy regulation) risks national and individual competitiveness.