Tencent Hunyuan-A13B Open-Source Review: 80B Total / 13B Active, GQA + FP8/INT4 Push MoE Inference Down to Consumer Hardware
Why "Active Parameters" Matter in MoE
Over the past year, frontier LLMs have split into two camps. One is the "dense + full activation" lane, exemplified by Llama 3 70B and Qwen2.5-72B. The other is the MoE lane, where the model routes only a small subset of experts per token, scaling capacity while keeping per-step compute manageable. DeepSeek-V3, Mixtral, and Qwen3-A22B all sit in this lane.
Tencent Hunyuan's latest open-source release, Hunyuan-A13B, is the newest entry on the MoE side. It carries 80B total parameters but activates only 13B per token, an order of magnitude smaller than Hunyuan-Large (389B total), and noticeably lighter than Qwen3-A22B (22B active). The positioning is clear: flagship-grade capability inside a mid-range enterprise compute budget.
The Core: Fine-Grained MoE + Dual-Mode Thinking + 256K Context
Hunyuan-A13B uses a fine-grained MoE architecture, slicing experts more aggressively than the older 8/16-expert coarse designs, which gives the router more flexibility. Combined with Grouped Query Attention (GQA), KV-cache memory pressure drops noticeably compared with multi-head attention, which directly helps long-context throughput.
The model natively supports a 256K context window. The README explicitly calls out stability on long-text tasks, which matters for codebase analysis, long-document QA, and long-horizon agent workloads.
The other interesting design choice is dual-mode reasoning. The model defaults to slow thinking (chain-of-thought), but users can prepend /think or /no_think to the prompt to force a mode, or pass enable_thinking=False to apply_chat_template to disable it outright. The direction is the same as Claude Opus 5's adaptive thinking, but Hunyuan exposes it as an explicit switch, which gives deployment-side more control.
Quantization: AngelSlim Covers BF16/FP8/INT4 With Negligible Loss
Quantization is the most interesting part of this release. Tencent did not just throw a GPTQ/INT4 checkpoint over the wall. They shipped the quantization tooling (AngelSlim) together with BF16, FP8 (static), and GPTQ-INT4 checkpoints, with first-class support for TensorRT-LLM, vLLM, and SGLang.
The accuracy drop is tiny. From the README's BF16 vs FP8/INT4 table:
| Bench | BF16 | FP8 | INT4 |
|---|---|---|---|
| AIME 2024 | 87.3 | 86.7 | 86.7 |
| GSM8k | 94.39 | 94.01 | 94.24 |
| BBH | 89.1 | 88.34 | 87.91 |
| DROP | 91.1 | 91.1 | 91.05 |
INT4 only loses 0.6 points on AIME 2024 versus BF16. For many private-deployment scenarios, this is the line between "deployable" and "not deployable".
Agent Benchmarks: Top of BFCL-v3 / τ-Bench / C3-Bench
The Instruct model's agent numbers are the most striking row in the comparison:
- BFCL-v3: 78.3 (Qwen3-A22B is 70.8, DeepSeek R1 is 56.9)
- ComplexFuncBench: 61.2 (Qwen3-A22B is 40.6, DeepSeek R1 is 41.1)
- C3-Bench: 63.5 (Qwen3-A22B is 51.7)
Math is also respectable: AIME 2024 87.3, AIME 2025 76.8, in the same league as DeepSeek R1 and slightly behind Qwen3-A22B. The trade-off Hunyuan-A13B offers on the agent utility axis, the dimension enterprises are actually paying for right now, is genuinely attractive.
Deployment: Three Backends, Docker Images Pre-Built
The deployment story is unusually polished:
- TensorRT-LLM: an official
hunyuaninfer/hunyuan-a13bDocker image with a one-command serve setup. - vLLM 0.8.5: dedicated image, CUDA 12.8+, TP=4 to spin up an OpenAI-compatible API.
- SGLang: same treatment, official image available.
The INT4 checkpoint should fit into roughly 40GB of weights for the 80B model, plus KV cache and batch overhead, so single 24GB cards are unrealistic, but dual 24GB cards or NVLink setups are workable. Compared with full BF16 on multi-card H100 clusters, this is the most pragmatic open-source MoE footprint available.
Commentary: Hunyuan Catches Up on "Practical Open Source"
Zooming out to the industry level:
- Fine-grained MoE is the consensus choice among Chinese open-source camps. Qwen3-A22B and DeepSeek V3 are both in this lane. Hunyuan-A13B's 80B/13B split is more conservative than A22B, clearly aimed at mid-size enterprise rooms with single 4-card or 8-card nodes.
- Quantization is the product. AngelSlim is not a bonus; it ships FP8/INT4 checkpoints as first-class artifacts. For private deployments, this effectively decouples "GPU budget" from "model capability".
- Agent benchmarks are becoming the hard procurement signal. Scores on BFCL-v3 / τ-Bench / C3-Bench are now the numbers enterprise buyers cite when justifying an open-source model purchase. Hunyuan-A13B's lead here will keep it on the shortlist for second-half 2026 enterprise agent projects.
So what? If you are picking an open-source MoE that can deploy locally, is forgiving on hardware, and can run agent workloads, Hunyuan-A13B is one of the most defensible PoC candidates right now. That said, 80B-total INT4 only means "runs". To actually carry long-horizon agents with 256K context, you still need to plan VRAM and throughput at A22B-equivalent scale. Do not be misled by the 13B-active headline.