{"version":"1.0","generated_at":"2026-10-10T00:39:27.199437","id":2092,"slug":"miles-blackwell-mxfp8-token-nvfp4","title":"Miles 在 Blackwell 架构上实现端到端 MXFP8 与逐 token NVFP4 强化学习方案","summary":"Miles团队在Blackwell架构上实现了端到端MXFP8和逐token NVFP4两种低精度强化学习方案。在8x B200上对Qwen3-30B-A3B的消融实验中，低精度配置与BF16的奖励曲线高度重合，且MXFP8和NVFP4减少了推理时间。","abstract":"Miles 在 Blackwell 架构上实现端到端 MXFP8 与逐 token NVFP4 强化学习方案 Miles团队在Blackwell架构上实现了端到端MXFP8和逐token NVFP4两种低精度强化学习方案。在8x B200上对Qwen3-30B-A3B的消融实验中，低精度配置与BF16的奖励曲线高度重合，且MXFP8和NVFP4减少了推理时间。 Miles团队在Blackwell架构上实现了端到端MXFP8和逐token NVFP4两种低精度强化学习方案。在8x B200上对Qwen3-30B-A3B的消融实验中，低精度配置与BF16的奖励曲线高度重合，且MXFP8和NVFP4减少了推理时间。 原贴提到：Miles 团队在 Blackwell 架构上实现了两种原生低精度强化学习方案：端到端 MXFP8 和 MoE 专家权重的逐 token NVF 来源：lmsys.org Miles团队在Blackwell架构上实现了两种原生低精度强化学习方案：端到端MXFP8和MoE专家权重的逐token NVFP4。在8x B200上对Qwen3-30B-A3B的消融实验中，BF16与所有五种低精度配置的原始奖励曲线高度重合，且MXFP8和NVFP4减少了推理时间。 信号解读： Miles团队在Blackwell架构上实现端到端MXFP8和逐token NVFP4的低精度强化学习方案，标志着低精度技术从推理向训练/微调环节延伸。在8x B200上对Qwen3-30B-A3B的实验中，低精度配置与BF16奖励曲线高度重合，说明低精度在RL场景下能保持性能，同时减少推理时间…","access_level":"member","access_label":"免费阅读","access_mode":"full","is_preview":false,"canonical_url":"https://opc.beizhux.com/content/2092/miles-blackwell-mxfp8-token-nvfp4","html_url":"https://opc.beizhux.com/content/2092/miles-blackwell-mxfp8-token-nvfp4","json_url":"https://opc.beizhux.com/content/2092/miles-blackwell-mxfp8-token-nvfp4.json","published_at":"2026-07-30T01:50:31","updated_at":"2026-10-09T23:54:53","category":{"slug":"ai-business","name":"AI小生意项目库"},"source":{"site":"lmsys.org","author":"LMSYS：Blog（Chatbot Arena 团队）","url":"https://www.lmsys.org/blog/2026-07-29-mxfp8-nvfp4-rl"},"tags":["AI","AIHOT 精选","Blackwell","Blackwell架构","Miles","MoE","MXFP8","NVFP4","低精度","低精度训练","强化学习"],"topics":[{"slug":"ai-side-business","name":"AI副业","url":"https://opc.beizhux.com/topics/ai-side-business","reason":"这篇内容命中「项目、小生意、变现」等主题信号。"},{"slug":"ai-super-individual","name":"AI超级个体","url":"https://opc.beizhux.com/topics/ai-super-individual","reason":"这篇内容命中「学习」等主题信号。"}],"keywords":["AI小生意项目库","AI副业","AI超级个体","模型","工作流","副业","创业","项目","小生意","变现","交付","内容"],"questions":[{"question":"Miles 在 Blackwell 架构上实现端到端 MXFP8 与逐 token NVFP4 强化学习方案主要讲什么？","answer":"Miles团队在Blackwell架构上实现了端到端MXFP8和逐token NVFP4两种低精度强化学习方案。在8x B200上对Qwen3-30B-A3B的消融实验中，低精度配置与BF16的奖励曲线高度重合，且MXFP8和NVFP4减少了推理时间。"},{"question":"这篇文章最值得关注的要点是什么？","answer":"Miles团队在Blackwell架构上实现了端到端MXFP8和逐token NVFP4两种低精度强化学习方案。在8x B200上对Qwen3-30B-A3B的消融实验中，低精度配置与BF16的奖励曲线高度重合，且MXFP8和NVFP4减…；原贴提到：Miles 团队在 Blackwell 架构上实现了两种原生低精度强化学习方案：端到端 MXFP8 和 MoE 专家权重的逐 token NVF；来源：lmsys.org"},{"question":"这篇文章和哪些AI专题相关？","answer":"它适合放在AI副业、AI超级个体专题里阅读。 关联原因：这篇内容命中「项目、小生意、变现」等主题信号。；这篇内容命中「学习」等主题信号。"},{"question":"阅读这篇文章建议先理解哪些关键词？","answer":"建议先理解模型、工作流、副业、创业、项目这些关键词，再结合正文判断工具、机会或风险是否值得进入自己的工作流。"}],"terms":[{"slug":"ai-model","name":"模型","definition":"在AI觉醒星球里，「模型」属于「AI工具」方向。围绕AI工具、模型能力、自动化流程和真实使用场景做整理，优先关注能提升效率、降低成本和创造新交付的工具。 不只看工具热度，更看它能不能进入真实流程。","topic_slug":"ai-tools","topic_name":"AI工具","topic_title":"AI工具：模型、插件、自动化与实战场景","topic_url":"https://opc.beizhux.com/topics/ai-tools","topic_path":"/topics/ai-tools","url":"https://opc.beizhux.com/glossary/ai-model","path":"/glossary/ai-model","json_url":"https://opc.beizhux.com/glossary/ai-model.json"},{"slug":"workflow","name":"工作流","definition":"在AI觉醒星球里，「工作流」属于「Agent工作流」方向。聚合Agent、AI编程、Claude Code、Cursor、工作流自动化等内容，帮助你把智能体能力落到真实任务链路。 把智能体当成任务系统，而不是聊天窗口。","topic_slug":"agent-workflow","topic_name":"Agent工作流","topic_title":"Agent工作流：智能体、AI编程与自动化协作","topic_url":"https://opc.beizhux.com/topics/agent-workflow","topic_path":"/topics/agent-workflow","url":"https://opc.beizhux.com/glossary/workflow","path":"/glossary/workflow","json_url":"https://opc.beizhux.com/glossary/workflow.json"},{"slug":"side-business","name":"副业","definition":"在AI觉醒星球里，「副业」属于「AI副业」方向。筛选AI副业、AI小生意、项目认知和可收费交付案例，关注普通人可以验证、复制和沉淀的方法。 从真实痛点出发，先做能收费的小验证。","topic_slug":"ai-side-business","topic_name":"AI副业","topic_title":"AI副业：小生意、项目机会与可收费交付","topic_url":"https://opc.beizhux.com/topics/ai-side-business","topic_path":"/topics/ai-side-business","url":"https://opc.beizhux.com/glossary/side-business","path":"/glossary/side-business","json_url":"https://opc.beizhux.com/glossary/side-business.json"},{"slug":"startup","name":"创业","definition":"在AI觉醒星球里，「创业」属于「AI副业」方向。筛选AI副业、AI小生意、项目认知和可收费交付案例，关注普通人可以验证、复制和沉淀的方法。 从真实痛点出发，先做能收费的小验证。","topic_slug":"ai-side-business","topic_name":"AI副业","topic_title":"AI副业：小生意、项目机会与可收费交付","topic_url":"https://opc.beizhux.com/topics/ai-side-business","topic_path":"/topics/ai-side-business","url":"https://opc.beizhux.com/glossary/startup","path":"/glossary/startup","json_url":"https://opc.beizhux.com/glossary/startup.json"},{"slug":"project","name":"项目","definition":"在AI觉醒星球里，「项目」属于「AI副业」方向。筛选AI副业、AI小生意、项目认知和可收费交付案例，关注普通人可以验证、复制和沉淀的方法。 从真实痛点出发，先做能收费的小验证。","topic_slug":"ai-side-business","topic_name":"AI副业","topic_title":"AI副业：小生意、项目机会与可收费交付","topic_url":"https://opc.beizhux.com/topics/ai-side-business","topic_path":"/topics/ai-side-business","url":"https://opc.beizhux.com/glossary/project","path":"/glossary/project","json_url":"https://opc.beizhux.com/glossary/project.json"},{"slug":"small-business","name":"小生意","definition":"在AI觉醒星球里，「小生意」属于「AI副业」方向。筛选AI副业、AI小生意、项目认知和可收费交付案例，关注普通人可以验证、复制和沉淀的方法。 从真实痛点出发，先做能收费的小验证。","topic_slug":"ai-side-business","topic_name":"AI副业","topic_title":"AI副业：小生意、项目机会与可收费交付","topic_url":"https://opc.beizhux.com/topics/ai-side-business","topic_path":"/topics/ai-side-business","url":"https://opc.beizhux.com/glossary/small-business","path":"/glossary/small-business","json_url":"https://opc.beizhux.com/glossary/small-business.json"},{"slug":"monetization","name":"变现","definition":"在AI觉醒星球里，「变现」属于「AI副业」方向。筛选AI副业、AI小生意、项目认知和可收费交付案例，关注普通人可以验证、复制和沉淀的方法。 从真实痛点出发，先做能收费的小验证。","topic_slug":"ai-side-business","topic_name":"AI副业","topic_title":"AI副业：小生意、项目机会与可收费交付","topic_url":"https://opc.beizhux.com/topics/ai-side-business","topic_path":"/topics/ai-side-business","url":"https://opc.beizhux.com/glossary/monetization","path":"/glossary/monetization","json_url":"https://opc.beizhux.com/glossary/monetization.json"},{"slug":"delivery","name":"交付","definition":"在AI觉醒星球里，「交付」属于「AI副业」方向。筛选AI副业、AI小生意、项目认知和可收费交付案例，关注普通人可以验证、复制和沉淀的方法。 从真实痛点出发，先做能收费的小验证。","topic_slug":"ai-side-business","topic_name":"AI副业","topic_title":"AI副业：小生意、项目机会与可收费交付","topic_url":"https://opc.beizhux.com/topics/ai-side-business","topic_path":"/topics/ai-side-business","url":"https://opc.beizhux.com/glossary/delivery","path":"/glossary/delivery","json_url":"https://opc.beizhux.com/glossary/delivery.json"}],"preview_hidden_sections":[],"related_items":[{"id":3305,"title":"Tibo 宣布发布 ChatGPT","url":"https://opc.beizhux.com/content/3305/tibo-chatgpt","summary":"Tibo 宣布 ChatGPT 发布，并给出访问地址 https://chatgpt.com/；推文本身仅含发布声明和链接，未提供更多功能或细节。","access_level":"member"},{"id":3306,"title":"OpenAI 年化营收被曝接近 500 亿美元，与此前预期差距约 200 亿美元","url":"https://opc.beizhux.com/content/3306/openai-500-200","summary":"金融时报报道称，OpenAI 向投资者披露截至 9 月底年化营收逼近 500 亿美元，低于外界估算的 700 亿美元约 200 亿美元；差异来自统计口径。消息发布后美股科技股下跌。","access_level":"member"},{"id":3307,"title":"Arena 完成 2 亿美元 B 轮融资，估值 31 亿美元并发布 Alignment Index","url":"https://opc.beizhux.com/content/3307/arena-2-b-31-alignment-index","summary":"Arena 完成 2 亿美元 B 轮融资，估值 31 亿美元，年化收入超 1 亿美元；同时发布基于真实智能体轨迹的 Alignment Index，从 Unauthorized Action、False Attribution 和 Deceptive Completion 三个信号评估 20 多个前沿模型。","access_level":"member"}],"usage_policy":{"summarizable":true,"preferred_url":"https://opc.beizhux.com/content/2092/miles-blackwell-mxfp8-token-nvfp4","private_fields_excluded":true}}