Nemotron 3 Ultra 技术报告解读:550B 规模的 NVFP4 预训练与多教师蒸馏

解读 NVIDIA Nemotron 3 Ultra:用 550B 参数与 20T token 验证 NVFP4 在极限规模下的训练稳定性,并通过多教师在线策略蒸馏 MOPD 把多个专用教师模型的能力压缩进单一学生模型。

Nemotron 3 Ultra —技术报告摘要

论文:Nemotron 3 Ultra: Open, Efficient Mixture-of-Experts Hybrid Mamba-Transformer Model for Agentic Reasoning (arXiv:2606.15007)
团队:NVIDIA
发布时间:2026.06
页数:65页


1. 预训练 Recipe

训练策略

项目详情
模型架构Hybrid Mamba-Attention MoE,550B总参/ 55B激活
层数108层,周期性Mamba-2+LatentMoE交错,少量Attention anchor
Model Dimension8192
注意力GQA:64Q-heads,2KV-heads, head dim128
Mamba-2State dim128, 8 groups, 256 heads, head dim64
LatentMoE512total experts/layer, top-22 activated, latent size 2048, expert hidden5120, shared expert10240
MTP2层shared-weight MTP,每层=1个attention layer+1个MoE layer;用于speculative decoding
训练精度NVFP4(目前最大规模NVFP4训练展示)
高精度保留最后15%网络(16层)+Mamba output+Latent projections+QKV+MTP+Embedding保持BF16/MXFP8
LR scheduleWarmup-Stable-Decay
上下文1M tokens

训练数据

项目详情
总量20T tokens,分两阶段
Phase 1(15T)多样性和广泛领域覆盖
Phase 2(5T)高质量数据精炼accuracy
合成数据(全部开源)Benchmark-oriented synthetic, fact-seeking (from FineWiki), moral scenarios, legal (California Code/NYCourts/eCFR/Case Law), 多种领域
训练健康监控5T/10T/16T checkpoints分支BF16 ablation 74B tokens,确认NVFP4无质量损失

Infra设计

项目详情
继承Nemotron 3 SuperNVFP4 GEMMs via Transformer Engine + cuBLAS
到date最大规模NVFP4训练550B参数×20T tokens

2. 后训练 Recipe (SFT)

训练策略

项目详情
两阶段SFTStage1: packed294,912 tokens, batch64, 204,800 samples, peak lr 1.5e-5 → Stage2: packed 515,000 tokens, 长上下文数据到512K, 19,200 samples, peak lr 1e-5
MTP保留SFT时保持shared-weight MTP objective(2层,aux-loss scale0.1)
数据覆盖Long-context, math/proof (1.8M tool-calling + 1.9M non-tool), science (web-search + Python traces), chat (multi-turn from GLM-5), agentic CLI, SWE, CUDA
Agentic数据含terminal-use, web browsing, code SWE, tool-calling等

3. RL Recipe

训练策略

项目详情
PipelineSFT → RLVR → MOPD Warmup → MOPD (xN cycles) → MTP Boosting
RLVRUnified verifiable-reward training:sharpens agentic, reasoning, chat, safety, instruction following, long-context
MOPD核心Multi-teacher On-Policy Distillation:多个specialized teacher模型提供token-level dense supervision
MOPD机制Student生成rollout →对应domain teacher对trajectory做token-level评分 → student学习teacher的preference
MOPD warmup一次性light SFT对齐student rollouts和teacher-supported分布
MOPD迭代Asynchronous on-policy distillation,2轮迭代(MOPD1→MOPD2),显著提升agentic benchmarks
Recovery率Terminal Bench 172.7%, SWE-Bench 88.1%, TauBench 90.3%(接近甚至超越specialized teacher)
MOPD vs logit matchingSampled-token objective一致优于full-distribution distillation(off-support states导致teacher logits poorly calibrated)
MTP BoostingHead-only KL distillation:MTP draft heads向backbone logits对齐,提升speculative decoding acceptance率
Reasoning budget control支持不同effort level

Teacher模型训练

项目详情
策略每domain训练一个specialized teacher(从student SFT checkpoint初始化 + 额外SFT/RL)
General reasoning teacher额外大规模SFT + RL on DeepSeek-V4-Pro生成的reasoning mixture
Agentic teachers针对terminal/browsing/SWE等domain独立训练

4. 关键亮点

  1. 最大规模NVFP4预训练:550B参数×20T tokens,验证NVFP4在extreme scale下的稳定性
  2. MOPD(Multi-teacher On-Policy Distillation):多specialized teacher→单一student,token-level dense supervision;多benchmarks上recovery >85%甚至超越teacher
  3. MTP Boosting:RL后通过KL distillation修复MTP heads,恢复speculative decoding质量
  4. 推理效率:vs GLM-5.1 5.9×,vs Kimi-K2.6 4.8×,vs Qwen3.5-397B 1.6×(8K/64K setting)
  5. 两阶段SFT + MTP auxiliary loss:保持长序列能力同时不损MTP
  6. MOPD局限性分析:对self-contained reasoning(如HLE)效果有限,因teacher优势来自off-policy数据exposure而非trajectory preference
本文结束 感谢您的阅读