[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"news-slug-nvidia-nemotron-3-super-120b-latent-moe":3,"news-related-267a9244-2ed7-4034-86cb-be4cbd196a08":36},{"id":4,"title":5,"summary":6,"content":6,"original_url":7,"source_id":8,"tags":9,"translations":23,"news_slug":29,"published_at":30,"created_at":31,"modified_at":32,"is_published":33,"publish_type":34,"image_url":13,"view_count":35},"267a9244-2ed7-4034-86cb-be4cbd196a08","NVIDIA 开源 Nemotron 3 Super：Latent MoE 如何让 120B 模型「省着跑」","NVIDIA 近日发布 Nemotron 3 Super（120B 总参数 \u002F 12B 激活参数），这是继去年 12 月 Nemotron 3 Nano 之后的第二代开源推理优化模型。不同于 Nano 的轻量定位，Super 直接面向多智能体（Multi-agent）场景——这类系统每个任务能生成普通对话 15 倍的 token 量，上下文爆炸和「思考税」是核心痛点。NVIDIA 的解法是一套从注意力到部署的完整架构创新。\n\nLatent MoE 把 4 倍专家塞进同一个 token。标准 MoE 将 token 直接路由到专家层，随着模型增大，路由层本身成为计算瓶颈。Nemotron 3 Super 引入了 Latent MoE：token 在进入专家前先被投影到一个低秩压缩空间，专家计算在更小的维度完成，结果再投影回完整维度。在同等推理成本下，可调用的专家数量提升 4 倍。更细粒度的专业化成为可能——比如一个专家专门处理 Python 语法，另一个专门处理 SQL 逻辑，各自仅在被需要时才激活。\n\nHybrid Mamba-Transformer 让线性复杂度和精确召回共存。Nemotron 3 Super 的 backbone 交替使用三类层：Mamba-2 层处理序列主体，引入状态空间模型（SSM）的线性时间复杂度，使得 100 万 token 的上下文在实际部署中可行；Transformer 注意力层在关键深度插入，保证模型在「大海捞针」类任务中维持高精度召回。这个组合解决了一个长期矛盾：Mamba 等 SSM 在长序列上计算友好，但精确关联召回弱；Transformer 注意力精确但 O(n²) 成本高，混合架构让两种能力各司其职。\n\nMulti-Token Prediction 一举两得。MTP 让模型在每个位置同时预测多个未来 token，训练阶段强制模型捕捉更长距离的依赖关系，产出的多 token 预测天然支持推测解码，减少自回归解码延迟。\n\n其他参数：原生 NVFP4 在 Blackwell 上推理速度提升 4 倍；超 120 万次环境 rollout 做多环境 RL 后训练；在 PinchBench 上得分 85.6%；权重、数据集、训练配方全部开源。\n\nNemotron 3 Super 真正值得关注的不是单一技术创新，而是「效率组合」的设计思路。Latent MoE 解决路由瓶颈，Hybrid backbone 解决上下文长度与召回质量的矛盾，MTP 同时改善训练和推理效率——这三层创新互相咬合，共同服务于同一个目标：让大参数模型在多智能体长程任务中跑得起、跑得快、跑得准。NVIDIA 的开源策略也在改变生态，Nemotron 3 Nano 开源后迅速成为 AI Agent 部署的标准基座，Super 以更大幅度提升效率，加上完整的训练配方，预计会复刻这一路径。","https:\u002F\u002Fdeveloper.nvidia.com\u002Fblog\u002Fintroducing-nemotron-3-super-an-open-hybrid-mamba-transformer-moe-for-agentic-reasoning\u002F","474eef8c-e0c3-46cf-adee-c089558220f9",[10,14,17,20],{"id":11,"name":12,"slug":12,"description":13,"color":13},"fca9258a-9430-455a-b95d-b9fae5e373a8","ai-inference",null,{"id":15,"name":16,"slug":16,"description":13,"color":13},"0ef8513a-0a26-42f0-b6f9-5b6dadded45c","efficiency",{"id":18,"name":19,"slug":19,"description":13,"color":13},"01598627-1ea6-4b27-a5d8-874971571a71","llm",{"id":21,"name":22,"slug":22,"description":13,"color":13},"b9bd9039-fcdb-41a8-b85b-fc1587def2b9","open-source",[24],{"id":25,"lang":26,"title":27,"summary":28,"content":13},"fa366ed1-b05e-4037-9498-18c6f60300bf","en","Nemotron 3 Super open-sourced: latent MoE runs 120B lean","NVIDIA recently released Nemotron 3 Super (120B total parameters \u002F 12B activated), the second-generation open-source reasoning-optimized model following December's Nemotron 3 Nano. Unlike Nano's lightweight positioning, Super directly targets the multi-agent scenario — these systems generate 15× the token volume of normal dialogue per task, with context explosion and \"thinking tax\" as core pain points. NVIDIA's solution is a complete architectural innovation from attention to deployment.\n\n**Latent MoE packs 4× more experts into the same token.** Standard MoE routes tokens directly to expert layers; as the model grows, the routing layer itself becomes the compute bottleneck. Nemotron 3 Super introduces Latent MoE: tokens are projected into a low-rank compressed space before entering experts, expert computation happens in a smaller dimension, and the result is projected back to the full dimension. Under the same inference cost, the number of callable experts increases 4×. Finer-grained specialization becomes possible — one expert handles Python syntax, another SQL logic, each activated only when needed.\n\n**Hybrid Mamba-Transformer makes linear complexity and precise recall coexist.** Nemotron 3 Super's backbone alternates three types of layers: Mamba-2 layers handle the main sequence, introducing the State Space Model's (SSM) linear-time complexity, making 1M-token context practically deployable; Transformer attention layers are inserted at critical depths to ensure high-precision recall on \"needle-in-a-haystack\" tasks. This combination resolves a long-standing contradiction: SSMs like Mamba are compute-friendly on long sequences but weak on precise-recall; Transformer attention is precise but O(n²) expensive. The hybrid architecture lets each capability play to its strength.\n\n**Multi-Token Prediction: two birds, one stone.** MTP has the model predict multiple future tokens at each position simultaneously; during training, the model is forced to capture longer-range dependencies, and the multi-token predictions naturally support speculative decoding, reducing autoregressive decoding latency.\n\nOther parameters: native NVFP4 delivers 4× inference speedup on Blackwell; over 1.2 million environment rollouts for multi-environment RL post-training; 85.6% on PinchBench; weights, datasets, and training recipe all open-sourced.\n\nWhat's really worth attention in Nemotron 3 Super isn't any single technical innovation, but the \"efficiency portfolio\" design philosophy. Latent MoE solves the routing bottleneck, the Hybrid backbone resolves the context-length vs recall-quality contradiction, and MTP improves both training and inference efficiency — these three layers of innovation interlock, all serving the same goal: making large-parameter models runnable, fast, and accurate on long-horizon multi-agent tasks. NVIDIA's open-source strategy is also reshaping the ecosystem: Nemotron 3 Nano became the standard base for AI Agent deployment shortly after open-sourcing. With Super's bigger efficiency boost and complete training recipe, this path is expected to repeat.","nvidia-nemotron-3-super-120b-latent-moe","2026-05-18T04:05:00Z","2026-05-18T04:05:22.847758Z","2026-08-19T02:08:40.142862Z",true,"agent",98,{"items":37},[38,43,48,53,58,63],{"id":39,"title":40,"news_slug":41,"published_at":42},"c6dc2edc-1a4a-46b9-85e6-f1c8ef32faa6","DeepSeek DSpark 跑进 Apple Silicon：mlx-dspark 给出首个原生 MLX 移植,逐字节保持原模型输出","mlx-dspark-apple-silicon","2026-07-04T12:00:00+00:00",{"id":44,"title":45,"news_slug":46,"published_at":47},"62e17707-e36f-45f6-8749-0d0370382cbd","llm-d：混合 GPU 集群 3-5 倍加速，KV Cache 感知路由","llm-d-mixed-gpu-kv-cache-aware-routing","2026-06-23T22:00:00+00:00",{"id":49,"title":50,"news_slug":51,"published_at":52},"68072ee1-fc37-4064-ab18-09550ae72d1b","GLM-5.3-Flash 把 320B MoE 跑在国产芯片上:Flash 价位和 $0.15 API 的混合注意力栈","glm-5-3-flash-chinese-chips-hybrid-attention","2026-08-27T03:00:00+00:00",{"id":54,"title":55,"news_slug":56,"published_at":57},"cac485ab-a429-4ebc-88c6-a1f924f978ff","AWS 开源 KeysAndValues:微调时就让模型学会“遗忘”,单张 A100 撑住 128K","aws-keysvalues-sparse-attention-finetuning","2026-08-26T05:20:00+00:00",{"id":59,"title":60,"news_slug":61,"published_at":62},"4aa9534a-778e-4cd7-8194-fdf3097249b8","OpenAI Jalapeño Hot Chips 实测:峰值每瓦 1.9×,延迟压到 1 秒","openai-jalapeno-hot-chips-benchmark-2026","2026-08-26T02:00:00+00:00",{"id":64,"title":65,"news_slug":66,"published_at":67},"48e1c261-a40a-4c71-9cba-450a459e6ad3","4-bit 模型反超全精度:QAH 把量化从性能税变成第二次蒸馏","quantization-aware-healing-hypernova-60b","2026-08-25T17:20:00+00:00"]