Hunyuan-A13B

2025-06-27 · Publisher : Tencent · Domain : Artificial Intelligence

Tencent opens Hunyuan-A13B, a mixture of experts (MoE) model with 80 billion parameters of which 13 billion are active, with fast or slow hybrid reasoning and a 256,000-token context, under the Tencent community license.

### Context In late June 2025, the open ecosystem is dominated by DeepSeek-R1 and the Qwen3 family, while OpenAI's closed reasoning models remain the reference in mathematics. Hunyuan-Large, Tencent's previous open model, serves as a point of comparison in the entry; Hunyuan-A13B aims at the opposite of a very large model: a compact model that can be used under resource constraints. ### What the model brings - Compact model: 13 billion active parameters out of 80 billion, which aims at performance close to larger models with a reduced inference cost. - Hybrid reasoning: slow thinking mode by default, which can be disabled with enable_thinking=False or the /no_think instruction, and forced with /think. - Native 256,000-token context, grouped-query attention (GQA) and several quantization formats (FP8 and GPTQ Int4 published). - Agent capabilities: Tencent highlights BFCL v3, τ-Bench and C3-Bench, where it claims leading results. - Two open evaluation sets released with the model: ArtifactsBench (visual and interactive code) and C3-Bench (agents). ### Architecture and specifications - Architecture: fine-grained MoE, 80 billion parameters in total, 13 billion active; 1 shared expert and 64 non-shared experts, 8 activated per pass. - Context: 256,000 tokens. Modalities: text. Languages: not detailed in the entry consulted. Training data and knowledge cutoff: not disclosed. - License: Tencent Hunyuan Community License. It excludes the territory of the European Union, the United Kingdom and South Korea (territory of application limited to the rest of the world) and requires requesting a license from Tencent beyond 100 million monthly active users. It is not a free license in the sense of Apache 2.0 or MIT. - Weights: Hunyuan-A13B-Instruct, FP8 version and GPTQ Int4 version on Hugging Face; pre-training model published. ### Cost and access No API pricing was consulted for this model. The vendor highlights running under resource constraints: the 13-billion-active-parameter model is designed to fit on few cards, but the entry does not set a minimum configuration. | Item | Information | | --- | --- | | Official API | Pricing not consulted | | Self-hosting | Minimum configuration not disclosed; FP8 and GPTQ Int4 available | | License | Tencent Hunyuan Community License (excluding EU, United Kingdom, South Korea) | ### Benchmarks | Benchmark | Hunyuan-A13B-Instruct | OpenAI-o1-1217 | DeepSeek R1 (open) | Qwen3-A22B (open) | | --- | --- | --- | --- | --- | | AIME 2024 | 87.3 | 74.3 | 79.8 | 85.7 | | AIME 2025 | 76.8 | 79.2 | 70.0 | 81.5 | | MATH | 94.3 | 96.4 | 94.9 | 94.0 | | GPQA-Diamond | 71.2 | 78.0 | 71.5 | 71.1 | | LiveCodeBench | 63.9 | 63.9 | 65.9 | 70.7 | | ZebraLogic | 84.7 | 81.0 | 78.7 | 80.3 | | IF-Eval | 84.7 | 91.8 | 88.3 | 83.4 | | BFCL v3 | 78.3 | 67.8 | 56.9 | 70.8 | | τ-Bench | 54.7 | 60.4 | 43.8 | 44.6 | | C3-Bench | 63.5 | 58.8 | 55.3 | 51.7 | Source: Hugging Face entry for Hunyuan-A13B-Instruct. Scores measured by Tencent, single table including OpenAI-o1-1217 (closed), DeepSeek R1 and Qwen3-A22B (open). ### Versus closed models Against OpenAI-o1-1217, the only closed model in the table, Hunyuan-A13B is ahead on AIME 2024 (87.3 versus 74.3), on the BFCL v3 function calls (78.3 versus 67.8) and on C3-Bench (63.5 versus 58.8). It is behind on AIME 2025 (76.8 versus 79.2), GPQA-Diamond (71.2 versus 78.0), instruction following IF-Eval (84.7 versus 91.8) and τ-Bench (54.7 versus 60.4); on LiveCodeBench the two are tied at 63.9. The comparison is limited to a closed model from late 2024 (the December 17, 2024 version, according to its name): it says nothing about the following generation. The price gap cannot be quantified for lack of consulted pricing. For an open or closed choice, the deciding point is the license: the exclusion of the European Union, the United Kingdom and South Korea prevents compliant use by a company established in Europe without a specific agreement from Tencent. ### Key takeaways Hunyuan-A13B shows that a MoE with 13 billion active parameters can match o1-1217 on several tests in the entry, with a 256,000-token context. Its license, which excludes the EU, the United Kingdom and South Korea, limits its use in Europe.

Tags : LLM, Open Source, Tencent, Hunyuan, MoE, Licence communautaire Tencent, Raisonnement hybride