Submit
TencentOpen-weight·131K context·389B params

Hunyuan-Largetencent/hunyuan-large

Hunyuan-Large is an open-source Mixture of Experts (MoE) model featuring 389 billion total parameters with 52 billion active parameters. It represents the largest open-source Transformer-based MoE model, with technical advantages including high-quality synthetic data training, KV cache compression via Grouped Query Attention and Cross-Layer Attention, expert-specific learning rate scaling, long-context processing up to 256K for pretraining and 128K for instruct variants, and extensive multilingual and multitask benchmarking. The model achieves state-of-the-art performance on standard benchmarks including MMLU, mathematical reasoning, and Chinese language tasks.

Pricing across providers

No pricing data yet. Verification in progress — check back soon.

Capabilities

  • instruction-following
  • reasoning
  • mathematics
  • code-generation
  • commonsense-understanding
  • long-context

Languages: en, zh

Benchmarks

  • HumanEvalinstruct · official · source
    9000.0%
  • HumanEvalinstruct model · official · source
    9000.0%
  • MMLUinstruct · official · source
    8990.0%
  • MMLUinstruct model · official · source
    8990.0%
  • MMLUpretrained · official · source
    8840.0%
  • MMLUpre-trained model · official · source
    8840.0%
  • HumanEvalpretrained · official · source
    7140.0%
  • HumanEvalpre-trained model · official · source
    7140.0%

Code samples

Technical specs

Context
131K
Max output
—
Parameters
389B
Release
2024-11-04
Training cutoff
—
License
—
Paper →Hugging Face →

Similar models

Compare with

  • Hunyuan-Large vs Llama 3.1 405B Instruct
    Comparison planned — not yet published
  • Hunyuan-Large vs Hunyuan Turbo
    Comparison planned — not yet published

Frequently asked

How much does Hunyuan-Large cost?+
No public pricing listed yet. Check back — we verify provider pricing on a 14-day cadence.
Is Hunyuan-Large open-source? Can I fine-tune it?+
Yes. Hunyuan-Large is open-weight under the unspecified license. Weights are available on Hugging Face for local inference, fine-tuning, and commercial use (see license for specific terms).
Is Hunyuan-Large OpenAI-compatible?+
Most listed hostings expose an OpenAI-compatible API, so you can point an existing openai SDK client at the Provider's base_url and use the Provider's model name. See the Code Samples above for a copy-pasteable example.
What's the maximum context window for Hunyuan-Large?+
The model supports up to 131,072 tokens of context (input + output). Some hosted versions may impose a smaller limit — check the "Context" column in the pricing matrix for each provider.