Hunyuan-A13B Instruct: Is the 2025 Tencent MoE Still Worth It?

Hunyuan-A13B was Tencent's mid-2025 open-weight mixture-of-experts release. It was open-sourced on 27 June 2025, it is small enough to self-host on a modest multi-GPU box, and it shipped with a switch between quick answers and step-by-step reasoning.

Sixteen months on, Tencent has released Hy3 and Hy4 preview, and the obvious question is whether A13B has any job left. It does, but a narrower one than its launch coverage suggested, and the deciding factor is more often the licence than the benchmarks.

What Tencent actually shipped

A13B is a fine-grained mixture-of-experts model with 80B total parameters and 13B active per token, using grouped-query attention. The model card does not publish layer or expert counts; those live in the technical report linked from it. If the total-versus-active distinction is new to you, mixture of experts explained covers why 13B active is the number that sets your per-token compute while 80B sets your memory bill.

Tencent also published quantised builds alongside the instruct model: an official FP8 checkpoint and a GPTQ Int4 one. The official serving examples cover vLLM, SGLang and TensorRT-LLM, with tensor parallelism of four in the vLLM and SGLang examples and two in the TensorRT-LLM one.

The vendor benchmark table for the instruct model is broad rather than spiky. A selection:

• Maths — AIME 2024 87.3, AIME 2025 76.8, MATH 94.3
• Science — GPQA-Diamond 71.2
• Coding — LiveCodeBench 63.9, FullStackBench 67.8
• Instruction following — IF-Eval 84.7
• Agent and tool use — BFCL v3 78.3, τ-Bench 54.7, C3-Bench 63.5

These are Tencent's numbers on Tencent's harness, from mid-2025. In Tencent's table they were competitive with larger models of the time, such as DeepSeek R1 and Qwen3-A22B. A year of releases has moved the field since, so do not read them against 2026 scores without running your own comparison.

Two thinking modes behind one switch

The feature that aged best is the dual-mode reasoning. Slow thinking is the default: the model produces an internal chain of thought before answering. Fast thinking skips it. You switch either by passing enable_thinking=False to the chat template, or by prefixing the prompt with /no_think. A /think prefix forces reasoning back on.

That matters for cost more than for quality. Reasoning tokens are billed as output, and output is the expensive side of the price sheet. A classification or extraction call that runs in slow mode can easily spend more tokens thinking than answering. The token cost of reasoning models walks through how quickly that adds up across a pipeline.

A practical default: run fast mode for anything with a checkable, short answer, and reserve slow mode for multi-step maths, planning and code where you have measured a difference.

The context window depends on who is serving it

The model card states a native context of 256K tokens. It also says the shipped config.json defaults to 32K to avoid out-of-memory errors, and that you must raise max_position_embeddings or pass a larger --max-model-len to vLLM to get the full window. Tencent's own guidance for 256K is bfloat16 weights and KV cache across four high-memory GPUs, with a warning that an FP8 KV cache may hurt output quality.

Hosted endpoints make their own choice. OpenRouter lists the model at 131,072 tokens of context. So the honest answer to "how much context does A13B have" is: 256K if you serve it yourself and pay for the memory, 32K if you run the defaults, and whatever your provider configured otherwise. Check the endpoint, not the model card.

Read the licence before the benchmarks

A13B ships under the Tencent Hunyuan Community License, not a standard open-source licence. Three clauses decide whether you can use it at all:

1. Territory. The agreement states that it does not apply in the European Union, the United Kingdom or South Korea. Use there is unlicensed.
2. Scale. If your products had more than 100 million monthly active users at the version's release date, you must request a separate licence, which Tencent may grant at its sole discretion.
3. Outputs. You may not use the model or its outputs to improve other AI models, other than Hunyuan derivatives. That rules out using it to generate synthetic training data for a different base model.

The 100M-user threshold affects almost nobody. The territory clause affects a lot of people, and the output clause catches teams that planned to distil. This is not legal advice; read the licence text yourself, and see open-weight model licences compared for how it sits against MIT and Apache releases.

The contrast with Tencent's newer work is sharp. The final Hy3 release moved to Apache 2.0, with none of these restrictions. If the licence is what blocked you from A13B, it no longer blocks you from the family.

Where it still makes sense in late 2026

On a hosted API the case is weak. OpenRouter lists A13B at $0.14 input and $0.57 output per million tokens at the time of writing, which is slightly more than it lists the newer, much larger Hy3 at. Paying more per token for an older, weaker model only makes sense if you have already validated A13B on a workload and do not want to re-test.

Self-hosting is where it keeps a role. An 80B model with official FP8 and Int4 builds fits on hardware that cannot hold Hy3's 295B. If you have a two- or four-GPU node, need data to stay on it, operate outside the excluded territories, and your tasks are within reach of a mid-2025 model, A13B is a reasonable choice. The Hy3 guide covers what the larger model needs to run.

A short evaluation before you commit

• Take twenty real tasks and run each in fast mode and slow mode. Record output tokens as well as accuracy; the gap tells you where reasoning pays.
• Run the same twenty through Hy3 on a hosted endpoint. If Hy3 wins clearly and costs the same, the only reason left for A13B is where it runs.
• If you need long context, test at the length you will actually use, on the exact serving configuration you will deploy.
• Confirm the licence position with whoever signs off on legal risk before any of this reaches production.