[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"news-slug-tencent-hunyuan-a13b-80b-13b-gqa-angelslim-moe":3,"news-related-d5203c93-2022-4769-a6d2-c7765ded2b40":41},{"id":4,"title":5,"summary":6,"content":7,"original_url":8,"source_id":9,"tags":10,"translations":27,"news_slug":34,"published_at":35,"created_at":36,"modified_at":37,"is_published":38,"publish_type":39,"image_url":14,"view_count":40},"d5203c93-2022-4769-a6d2-c7765ded2b40","腾讯混元 Hunyuan-A13B 开源实测:80B 总参 \u002F 13B 激活,GQA + FP8\u002FINT4 把 MoE 推理门槛打到消费卡","腾讯混元在 GitHub 开源 Hunyuan-A13B:总参数 80B、激活 13B 的细粒度 MoE,原生 256K 上下文、混合快慢思考;用 Grouped Query Attention 与自研 AngelSlim 量化工具体系覆盖 BF16\u002FFP8\u002FINT4 部署,能在 TensorRT-LLM、vLLM、SGLang 上直接拉起,BF16 与 FP8\u002FINT4 量化后性能几乎不掉。","# 腾讯混元 Hunyuan-A13B 开源实测:80B 总参 \u002F 13B 激活,GQA + FP8\u002FINT4 把 MoE 推理门槛打到消费卡\n\n## 背景:MoE 模型的\"激活参数\"为何重要\n\n过去一年,主流大模型明显分成了两条路线。一条是\"堆总参数 + 全激活\"的密集型路线,代表是 Llama 3 70B、Qwen2.5-72B 这种全量激活的稠密模型;另一条是\"总参数大、激活参数小\"的 MoE 路线,通过路由只激活一小撮专家,在保持单次推理算力可控的前提下放大模型容量。DeepSeek-V3、Mixtral、Qwen3-A22B 都属于后者。\n\n腾讯混元最近在 GitHub 开源的 Hunyuan-A13B,是这条路线的最新一站。它的总参数是 80B,但每次推理只激活 13B,比 Hunyuan-Large(389B 总参)小一个量级,激活参数也比 Qwen3-A22B(22B 激活)更小,瞄准的就是\"在企业算力预算内塞下旗舰能力\"这个长期痛点。\n\n## 核心:细粒度 MoE + 双模思考 + 256K 上下文\n\nHunyuan-A13B 是细粒度(fine-grained)MoE,把专家切得更细,路由更灵活,这跟之前业内常见的 8\u002F16 粗粒度专家结构有明显区别。配合 Grouped Query Attention(GQA),KV Cache 占用比 Multi-Head Attention 低一截,长上下文场景下显存压力直接下降一档。\n\n模型原生支持 256K 上下文,官方 README 强调在长文本任务上表现稳定,这对代码库分析、长文档问答、Agent 长链路任务都直接有用。\n\n另一个有意思的设计是双模思考:模型默认走慢思考(slow thinking \u002F CoT),但用户可以在 prompt 前加 \u002Fthink 或 \u002Fno_think 强制切换,或者调用 apply_chat_template 时把 enable_thinking 设为 False 直接关掉。这跟 Claude Opus 5 那条 adaptive thinking 路线一脉相承,只是 Hunyuan 给的是显式开关,部署端可控性更强。\n\n## 量化体系:AngelSlim 一把梭,FP8\u002FINT4 几乎不掉点\n\n量化是这次开源最值得展开的部分。腾讯没有像别家那样丢一个 GPTQ\u002FINT4 checkpoint 就完事,而是把量化工具体系 AngelSlim 和量化模型(BF16\u002FFP8\u002FINT4 checkpoint)一起放出,覆盖 BF16、FP8(static)、GPTQ-INT4 三档,直接对接 TensorRT-LLM、vLLM、SGLang 三大推理框架。\n\n性能损失几乎可以忽略。README 给的对照表(基模型 vs FP8\u002FINT4):\n\n| Bench | BF16 | FP8 | INT4 |\n|---|---|---|---|\n| AIME 2024 | 87.3 | 86.7 | 86.7 |\n| GSM8k | 94.39 | 94.01 | 94.24 |\n| BBH | 89.1 | 88.34 | 87.91 |\n| DROP | 91.1 | 91.1 | 91.05 |\n\nAIME 2024 这种高难度数学基准,INT4 跟 BF16 只差 0.6 分,在很多企业私有部署场景里,这就是\"能跑\"和\"跑不起\"的区别。\n\n## Agent 能力:BFCL-v3 \u002F τ-Bench \u002F C3-Bench 拿第一档\n\nInstruct 版本在 Agent 评测上是这张表最有信号的一行:\n\n- BFCL-v3:78.3(Qwen3-A22B 是 70.8,DeepSeek R1 是 56.9)\n- ComplexFuncBench:61.2(Qwen3-A22B 是 40.6,DeepSeek R1 是 41.1)\n- C3-Bench:63.5(Qwen3-A22B 是 51.7)\n\n数学上 Hunyuan-A13B 也不弱:AIME 2024 87.3、AIME 2025 76.8,跟 DeepSeek R1 一个量级,小幅落后 Qwen3-A22B。综合来看,它在 Agent 实用性这个当下最被企业买单的方向上,给出的权衡确实够厚道。\n\n## 部署:三个推理后端,Docker 镜像都打包好了\n\n部署侧的友好度是这个 repo 让人省心的部分:\n\n- TensorRT-LLM:官方给了 hunyuaninfer\u002Fhunyuan-a13b 的 Docker 镜像,起服务一条命令\n- vLLM 0.8.5:有专门的镜像,cuda 12.8 起步,TP=4 就能拉起 OpenAI 兼容 API\n- SGLang:同样有官方镜像\n\nINT4 量化版本理论上能塞进单卡 24GB 显存(80B 总参 INT4 约 40GB,实际推理时再加 KV 缓存和 batch 开销可能需要双卡或者 NVLink),但相对全量 BF16 需要多卡 H100,这已经是开源 MoE 落地最务实的姿态了。\n\n## 个人评论:Hunyuan 在\"实用开源\"这个赛道上追平了\n\n把视角拉到行业层面:\n\n1. 细粒度 MoE 是中国开源阵营的共同选择。Qwen3-A22B、DeepSeek V3 都是这条路。Hunyuan-A13B 选 80B\u002F13B 这个档位,比 A22B 更克制,显然瞄准的是中等企业机房、单台 8 卡或 4 卡节点的预算线。\n2. 量化即产品。AngelSlim 这套工具不是额外福利,而是把 FP8\u002FINT4 checkpoint 当作一等公民交付。对私有部署来说,这相当于把\"买卡预算\"和\"模型能力\"解耦了。\n3. Agent 评测权重在变化。BFCL-v3 \u002F τ-Bench \u002F C3-Bench 这类 agentic benchmark 的分数,正在变成企业采购开源模型的硬指标。Hunyuan-A13B 在这一档的领先,会让它在 2026 下半年的企业 Agent 选型里被反复点名。\n\n所以呢:如果你正在选一个能本地部署、不挑卡、能跑 Agent 的开源 MoE,Hunyuan-A13B 是当下最值得放进 PoC 列表的候选之一。但要注意,80B 总参的 INT4 量化只是\"能跑\",真要做 Agent 长链路 + 256K 上下文,显存和吞吐还是要按 A22B 同等量级去规划,别被 13B 激活的纸面数字误导。","https:\u002F\u002Fhuggingface.co\u002Ftencent\u002FHunyuan-A13B-Instruct","998df6db-96e6-4b8e-8be1-cfa00a6cd177",[11,15,18,21,24],{"id":12,"name":13,"slug":13,"description":14,"color":14},"a8002d98-9df1-4ab9-94d4-a7625af634c4","china-ai",null,{"id":16,"name":17,"slug":17,"description":14,"color":14},"0a93ec8e-ea39-4693-81de-563ca8c173f7","inference",{"id":19,"name":20,"slug":20,"description":14,"color":14},"01598627-1ea6-4b27-a5d8-874971571a71","llm",{"id":22,"name":23,"slug":23,"description":14,"color":14},"b9bd9039-fcdb-41a8-b85b-fc1587def2b9","open-source",{"id":25,"name":26,"slug":26,"description":14,"color":14},"b49648f9-963e-4082-8684-3d085b7358fe","quantization",[28],{"id":29,"lang":30,"title":31,"summary":32,"content":33},"8e850fd9-bd34-495a-884f-abfd1deadc69","en","Hunyuan-A13B tested: 80B\u002F13B MoE on consumer GPUs","Tencent Hunyuan open-sourced Hunyuan-A13B on GitHub: a fine-grained MoE with 80B total and 13B active parameters, native 256K context, and dual fast\u002Fslow thinking modes. Backed by Grouped Query Attention and the in-house AngelSlim quantization stack (BF16\u002FFP8\u002FINT4), the model deploys cleanly on TensorRT-LLM, vLLM, and SGLang with negligible accuracy loss after quantization.","# Tencent Hunyuan-A13B Open-Source Review: 80B Total \u002F 13B Active, GQA + FP8\u002FINT4 Push MoE Inference Down to Consumer Hardware\n\n## Why \"Active Parameters\" Matter in MoE\n\nOver 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.\n\nTencent 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.\n\n## The Core: Fine-Grained MoE + Dual-Mode Thinking + 256K Context\n\nHunyuan-A13B uses a fine-grained MoE architecture, slicing experts more aggressively than the older 8\u002F16-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.\n\nThe 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.\n\nThe other interesting design choice is dual-mode reasoning. The model defaults to slow thinking (chain-of-thought), but users can prepend `\u002Fthink` or `\u002Fno_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.\n\n## Quantization: AngelSlim Covers BF16\u002FFP8\u002FINT4 With Negligible Loss\n\nQuantization is the most interesting part of this release. Tencent did not just throw a GPTQ\u002FINT4 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.\n\nThe accuracy drop is tiny. From the README's BF16 vs FP8\u002FINT4 table:\n\n| Bench | BF16 | FP8 | INT4 |\n|---|---|---|---|\n| AIME 2024 | 87.3 | 86.7 | 86.7 |\n| GSM8k | 94.39 | 94.01 | 94.24 |\n| BBH | 89.1 | 88.34 | 87.91 |\n| DROP | 91.1 | 91.1 | 91.05 |\n\nINT4 only loses 0.6 points on AIME 2024 versus BF16. For many private-deployment scenarios, this is the line between \"deployable\" and \"not deployable\".\n\n## Agent Benchmarks: Top of BFCL-v3 \u002F τ-Bench \u002F C3-Bench\n\nThe Instruct model's agent numbers are the most striking row in the comparison:\n\n- BFCL-v3: 78.3 (Qwen3-A22B is 70.8, DeepSeek R1 is 56.9)\n- ComplexFuncBench: 61.2 (Qwen3-A22B is 40.6, DeepSeek R1 is 41.1)\n- C3-Bench: 63.5 (Qwen3-A22B is 51.7)\n\nMath 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.\n\n## Deployment: Three Backends, Docker Images Pre-Built\n\nThe deployment story is unusually polished:\n\n- TensorRT-LLM: an official `hunyuaninfer\u002Fhunyuan-a13b` Docker image with a one-command serve setup.\n- vLLM 0.8.5: dedicated image, CUDA 12.8+, TP=4 to spin up an OpenAI-compatible API.\n- SGLang: same treatment, official image available.\n\nThe 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.\n\n## Commentary: Hunyuan Catches Up on \"Practical Open Source\"\n\nZooming out to the industry level:\n\n1. 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\u002F13B split is more conservative than A22B, clearly aimed at mid-size enterprise rooms with single 4-card or 8-card nodes.\n2. Quantization is the product. AngelSlim is not a bonus; it ships FP8\u002FINT4 checkpoints as first-class artifacts. For private deployments, this effectively decouples \"GPU budget\" from \"model capability\".\n3. Agent benchmarks are becoming the hard procurement signal. Scores on BFCL-v3 \u002F τ-Bench \u002F 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.\n\nSo 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.","tencent-hunyuan-a13b-80b-13b-gqa-angelslim-moe","2026-07-30T06:00:00Z","2026-07-30T02:06:40.993671Z","2026-07-30T02:06:40.993680Z",true,"agent",93,{"items":42},[43,48,53,58,63,68],{"id":44,"title":45,"news_slug":46,"published_at":47},"988bbfb8-672a-4c6c-98f7-3a170b6bd8b3","Macaw 把 LFM2.5 装进 1.5GB:4-bit 端侧 LLM 跑 Mac 控制工具链","macaw-lfm25-15gb-edge-mac-agent","2026-08-24T06:00:00+00:00",{"id":49,"title":50,"news_slug":51,"published_at":52},"1abe59b8-844d-4bb1-bc27-cfe28099d101","Anemll\u002FFlash-iOS：把 400B MoE 大模型塞进 iPhone 的端侧实验","anemll-flash-ios-400b-moe-iphone-edge","2026-06-07T12:00:00+00:00",{"id":54,"title":55,"news_slug":56,"published_at":57},"c2e18ecb-00b3-47ee-935a-f0e3cd0dee4a","Holo3.1 把 Computer Use Agent 拉进本地：FP8\u002FNVFP4\u002FQ4 三种量化让消费级 GPU 跑得动","holo3-1-h-company-computer-use-quant","2026-06-07T06:00:00+00:00",{"id":59,"title":60,"news_slug":61,"published_at":62},"77cd56b5-87cf-4ffe-9309-a85ae5d81da9","国家超算互联网上线DeepSeek-V4限时免费服务：百万Token上下文国产大模型开放新入口","china-supercomputing-deepseek-v4-free-1m-context","2026-04-26T13:00:00+00:00",{"id":64,"title":65,"news_slug":66,"published_at":67},"f6e4aab0-7693-4c2c-bb66-c1641fc2cc3e","Ox Alpha 谜底揭晓:智谱 GLM-5.3-Flash,MIT 开源 320B MoE","ox-alpha-glm-5-3-flash-reveal","2026-08-27T13:30:00+00:00",{"id":69,"title":70,"news_slug":71,"published_at":72},"804ab59a-a8d6-4b61-bf74-8f6f2bdae83c","智谱把 Flash 做成一件正经事:一次说清 GLM-5.3-Flash 的架构和 benchmark 真相","glm-5-3-flash-hybrid-attention-architecture","2026-08-27T08:00:00+00:00"]