Tencent’s Hunyuan staff launched it Hunyuan-A13ba brand new open supply main language mannequin constructed on sparse Combination (MOE) Structure. The mannequin consists of a complete of 80 billion parameters, however solely 13 billion are energetic throughout inference, offering a really environment friendly stability between efficiency and computational price. It helps Grouped Question Notes (GQA), 256K context sizea Twin-mode inference framework This switches between quick and gradual considering.
Designed for environment friendly deployment and sturdy inference, the Hunyuan-A13B BFCL-V3, τ bench, C3 benchand ComplexFuncbenchtypically outperforms the bigger fashions in instrument calls and lengthy context eventualities.
Structure: Sparse MOE with 13B energetic parameters
On the coronary heart of this, the Hunyuan-A13B is 1 Shared Knowledgeable and 64 non-shared specialistsand Eight specialists activated per ahead go. Supported by scaling experiments, this structure ensures efficiency consistency whereas protecting inference prices low. The mannequin contains 32 layers, makes use of Swiggle Activation, which is a vocabulary dimension of 128K, integrates GQA for elevated reminiscence effectivity throughout lengthy context inference.
The mannequin’s MOE setup is mixed with optimized ones Coaching Curriculum: Pre-deletion stage of the 20t token, adopted by quick annealing and lengthy context adaptation. This last section first scales the context window to 32K after which to 256K tokens utilizing NTK-enabled place encoding, making certain steady efficiency with massive sequence lengths.
Twin-mode reasoning: Quick and gradual considering
The excellent characteristic of the Hunyuan-A13B is itself Twin Mode Chain of Shart (COT) capacity. Helps each degradation I feel rapidly Routine Question Modes and Extra Elaborate Modes I am considering slowly Multi-step inference mode. These modes are managed by way of a easy tag system. /no assume For fast reasoning /assume For reflexive reasoning. This flexibility permits customers to adapt computational prices to process complexity.
Submit-training: Reinforcement studying with task-specific reward fashions
Included within the Hunyuan-A13B post-training pipeline Multi-stage Monitoring Effective Tuning (SFT) and Reinforcement Studying (RL) throughout each inference-specific and common duties. It’s included into the RL stage Outcomes-based rewards and Device-specific suggestionsConsists of checking the code sandbox execution setting and agent rule-based.
In the course of the agent coaching section, groups synthesize and generate a wide range of instrument utilization eventualities with the roles of planners, checkers and instruments. 20,000 format mixtures. This enhances the power of Hunyuan-A13B to carry out actual workflows equivalent to spreadsheet processing, info looking out, and structured inference.
Ranking: innovative agent efficiency
Hunyuan-A13b Present Robust benchmark outcomes Past the various NLP duties:
- Above Arithmetic, cmathand GPQAit scores greater than the bigger, extra dense and MOE fashions.
- That is past QWEN3-A22B and Deepseek R1 in Logical reasoning (BBH: 89.1; Zebralogic: 84.7).
- In coding, I’ve its personal retention at 83.9 on MBPP and 69.3 on Multipl-E.
- for Agent Processit continues BFCL-V3 (78.3) and ComplexFuncbench (61.2)to confirm its instrument utilization capabilities.
Understanding lengthy contexts is one other spotlight. Above Penguin Crawlit scores 87.7 – shy within the Gemini 2.5 Professional. Above rulerit maintains excessive efficiency (73.9) even 64K-128K contextoutperforms bigger fashions such because the QWEN3-A22B and DeepSeek R1 in context resilience.
Optimizing and increasing inference
Hunyuan-A13B is totally built-in with a common reasoning framework like vllm, sglangand Tensort-llm. Helps the next precision codecs: W16A16, W8A8and KV Cache FP8With options like Computerized Prefix Caching and Chunk Prill. I will obtain it 1981.99 tokens/sec Throughput of 32 batch inputs (2048 inputs, 14336 output size) makes them sensible for real-time purposes.
Open Supply and Business Relationship
Accessible at Hugging my face and githubHunyuan-A13B is launched with an appropriate open supply license. It’s designed for environment friendly analysis and manufacturing use, particularly in latency sensitivity-sensitive environments and lengthy contextual duties.
Mix it MOE Scalability, Agent’s reasoningand Open Supply AccessibilityTencent’s Hunyuan-A13B affords a compelling various to heavyweight LLMS, permitting for wider experimentation and deployment with out sacrificing capabilities.
Please test paper. All credit for this examine will probably be despatched to researchers on this venture. Additionally, please be at liberty to observe us Twitter And do not forget to hitch us 100k+ ml subreddit And subscribe Our Newsletter.
Asif Razzaq is CEO of Marktechpost Media Inc.. As a visionary entrepreneur and engineer, ASIF is dedicated to leveraging the probabilities of synthetic intelligence for social advantages. His newest efforts are the launch of MarkTechPost, a synthetic intelligence media platform. That is distinguished by its detailed protection of machine studying and deep studying information, and is straightforward to grasp by a technically sound and vast viewers. The platform has over 2 million views every month, indicating its reputation amongst viewers.
