{ "version": 1, "event_id": "evt_8d32783c978fc24d", "url": "https://xiyu.news/events/evt_8d32783c978fc24d/", "json": "https://xiyu.news/api/events/evt_8d32783c978fc24d.json", "type": "other", "status": "monitoring", "category": "technology", "title": { "zh": "Reflection 推出 Beam：一款以更低算力成本对标中国模型的开源权重 AI 模型", "en": "Reflection debuts Beam, an open-weight AI model to rival Chinese models at lower compute cost" }, "current_state": { "zh": "Beam 1.0 的完整权重已于 2026 年 10 月 16 日前后在 Hugging Face 上以 Apache 2.0 协议发布，提供 BF16、FP8 和 NVFP4 版本，支持 100 万 token 上下文。Reflection 同时开源了 Mirror CLI 编程智能体框架和 DFlash2 推测解码加速模型，并与 vLLM 合作优化推理性能。Artificial Analysis 独立评测给 Beam 打 30 分，在美国开放权重模型中排名第二，但落后于 DeepSeek、GLM 等中国模型。发布获 40 多家合作伙伴支持。英伟达进一步注资或收购 Reflection 的谈判仍在进行中。", "en": "Beam 1.0's full weights were released on Hugging Face under Apache 2.0 around October 16, 2026, with BF16, FP8 and NVFP4 versions and 1M-token context. Reflection also open-sourced the Mirror CLI coding agent framework and DFlash2 speculative decoding accelerator, and worked with vLLM to optimize inference. Independent Artificial Analysis evaluation gives Beam a score of 30, ranking second among US open-weight models but behind Chinese models such as DeepSeek and GLM. The release has support from over 40 partners. Nvidia's reported talks to further invest in or acquire Reflection are ongoing." }, "first_seen_at": "2026-10-05T23:27:29.704060+00:00", "last_updated_at": "2026-10-16T16:51:55.152680+00:00", "last_material_change_at": "2026-10-16T16:51:55.152680+00:00", "confidence": 0.75, "updates_count": 3, "sources_count": 3, "entities": [ "beam", "chinese", "meta", "reflection" ], "identifiers": [], "topics": [ "agent-framework", "ai-model-release", "ai-models", "benchmarks", "enterprise-ai", "inference-optimization", "mixture-of-experts", "moe", "open-weights", "reflection", "reinforcement-learning", "sovereign-ai" ], "updates": [ { "update_id": "upd_7056aa3928524fbe", "event_id": "evt_8d32783c978fc24d", "occurred_at": "2026-10-05T19:33:53Z", "published_at": "2026-10-05T19:33:53Z", "first_seen_at": "2026-10-05T23:27:29.704060Z", "time_precision": "published", "update_type": "initial", "material_change": true, "title_zh": "Reflection 推出 Beam：一款以更低算力成本对标中国模型的开源权重 AI 模型", "title_en": "Reflection debuts Beam, an open-weight AI model to rival Chinese models at lower compute cost", "what_changed_zh": "2026 年 10 月 5 日（周一），Reflection AI 发布了其首个前沿开源权重模型 Beam。据该公司博客介绍，Beam 是一个稀疏混合专家（MoE）模型，总参数 5010 亿、激活参数 230 亿，面向编程、推理与智能体（agentic）任务。\n\n此次发布使一家美国实验室在编程及相关任务上，直接与 DeepSeek、月之暗面 Kimi 等低成本中国开源权重模型展开竞争，而这一领域此前由中国模型的成本水平树立标杆。\n\nReflection 由英伟达（Nvidia）投资，并以“AI 工厂”（AI factories）产品向企业和主权国家推销 Beam 及后续模型，允许机构用自有专有数据训练这些模型。目前公开报道中未包含独立的基准测试结果或授权条款。", "what_changed_en": "Reflection AI unveiled Beam on Monday, October 5, 2026, its first frontier open-weight model. According to the company's blog, Beam is a sparse Mixture-of-Experts model with 501 billion total parameters and 23 billion active, built for coding, reasoning and agentic workloads.\n\nThe launch puts a US lab in direct competition with lower-cost Chinese open-weight models such as DeepSeek and Moonshot AI's Kimi in coding and related tasks, a segment where Chinese releases have set the cost benchmark.\n\nReflection is Nvidia-backed and is pitching Beam and future models at enterprises and sovereign nations through an \"AI factories\" offering that would let institutions train the models on their own proprietary data. The available coverage does not include independent benchmark results or licensing terms.", "current_state_zh": "2026 年 10 月 5 日（周一），Reflection AI 发布了其首个前沿开源权重模型 Beam。据该公司博客介绍，Beam 是一个稀疏混合专家（MoE）模型，总参数 5010 亿、激活参数 230 亿，面向编程、推理与智能体（agentic）任务。\n\n此次发布使一家美国实验室在编程及相关任务上，直接与 DeepSeek、月之暗面 Kimi 等低成本中国开源权重模型展开竞争，而这一领域此前由中国模型的成本水平树立标杆。\n\nReflection 由英伟达（Nvidia）投资，并以“AI 工厂”（AI factories）产品向企业和主权国家推销 Beam 及后续模型，允许机构用自有专有数据训练这些模型。目前公开报道中未包含独立的基准测试结果或授权条款。", "current_state_en": "Reflection AI unveiled Beam on Monday, October 5, 2026, its first frontier open-weight model. According to the company's blog, Beam is a sparse Mixture-of-Experts model with 501 billion total parameters and 23 billion active, built for coding, reasoning and agentic workloads.\n\nThe launch puts a US lab in direct competition with lower-cost Chinese open-weight models such as DeepSeek and Moonshot AI's Kimi in coding and related tasks, a segment where Chinese releases have set the cost benchmark.\n\nReflection is Nvidia-backed and is pitching Beam and future models at enterprises and sovereign nations through an \"AI factories\" offering that would let institutions train the models on their own proprietary data. The available coverage does not include independent benchmark results or licensing terms.", "detailed_summary_zh": "2026 年 10 月 5 日（周一），Reflection AI 发布了其首个前沿开源权重模型 Beam。据该公司博客介绍，Beam 是一个稀疏混合专家（MoE）模型，总参数 5010 亿、激活参数 230 亿，面向编程、推理与智能体（agentic）任务。\n\n此次发布使一家美国实验室在编程及相关任务上，直接与 DeepSeek、月之暗面 Kimi 等低成本中国开源权重模型展开竞争，而这一领域此前由中国模型的成本水平树立标杆。\n\nReflection 由英伟达（Nvidia）投资，并以“AI 工厂”（AI factories）产品向企业和主权国家推销 Beam 及后续模型，允许机构用自有专有数据训练这些模型。目前公开报道中未包含独立的基准测试结果或授权条款。", "detailed_summary_en": "Reflection AI unveiled Beam on Monday, October 5, 2026, its first frontier open-weight model. According to the company's blog, Beam is a sparse Mixture-of-Experts model with 501 billion total parameters and 23 billion active, built for coding, reasoning and agentic workloads.\n\nThe launch puts a US lab in direct competition with lower-cost Chinese open-weight models such as DeepSeek and Moonshot AI's Kimi in coding and related tasks, a segment where Chinese releases have set the cost benchmark.\n\nReflection is Nvidia-backed and is pitching Beam and future models at enterprises and sovereign nations through an \"AI factories\" offering that would let institutions train the models on their own proprietary data. The available coverage does not include independent benchmark results or licensing terms.", "background_zh": "开源权重指已训练 AI 模型公开释放的学习参数，是否允许修改、微调或再分发取决于具体许可证。DeepSeek、阿里云、月之暗面（Moonshot AI）和 Z.ai 等中国企业普遍以 Apache 或 MIT 等宽松许可发布开源权重模型，而美国大型实验室更倾向于闭源发布。截至 2026 年 7 月，规模最大的开源权重前沿模型是月之暗面的 Kimi K3（2.8 万亿参数）和阿里云的 Qwen3.8（2.4 万亿参数）。", "background_en": "Open weights are the publicly released learned parameters of a trained AI model; permission to modify, fine-tune or redistribute them depends on the license. Chinese companies including DeepSeek, Alibaba Cloud, Moonshot AI and Z.ai have generally released open-weight models under permissive licenses such as Apache or MIT, while large US labs have favored proprietary releases. As of July 2026, the largest open-weight frontier models were Moonshot AI's Kimi K3 at 2.8 trillion parameters and Alibaba Cloud's Qwen3.8 at 2.4 trillion.", "community_discussion_zh": "", "community_discussion_en": "", "market_impact_zh": "开源权重模型的发布会把部分模型部署转向自托管和本地司法辖区内的“AI 工厂”模式，这一渠道影响的是 GPU 与基础设施需求、推理托管和 API 定价，而非直接作用于交易市场。Reflection 获得英伟达投资，并以主权国家与企业为销售对象，因此与基础设施和算力需求这一议题相关联。", "market_impact_en": "Open-weight releases shift part of model deployment toward self-hosting and in-jurisdiction \"AI factory\" setups, a channel that touches GPU and infrastructure demand, inference hosting and API pricing rather than trading directly. Reflection's Nvidia backing and its sovereign-enterprise go-to-market place it in that infrastructure and compute-demand discussion.", "importance_score": 7.0, "references": [ { "url": "https://reflection.ai/blog/introducing-beam", "title": "Introducing Beam: Reflection’s 501B open-weight model" }, { "url": "https://www.reuters.com/technology/nvidia-backed-reflection-unveils-first-ai-model-take-chinese-open-models-2026-10-05/", "title": "Nvidia-backed Reflection unveils first AI model to take on ..." }, { "url": "https://techcrunch.com/2026/10/05/reflection-debuts-beam-a-open-weight-ai-model-to-rival-chinese-models-at-lower-compute-cost/", "title": "Reflection debuts Beam, an open-weight AI model to rival ..." } ], "confidence": 0.75, "story_ids": [ "rss:techcrunch.com_category_artificial-intelligence_feed_:2fd90060880f9693" ], "sources": [ { "url": "https://techcrunch.com/2026/10/05/reflection-debuts-beam-a-open-weight-ai-model-to-rival-chinese-models-at-lower-compute-cost/", "label": "TechCrunch AI", "source_type": "rss", "official": false } ] }, { "update_id": "upd_a68e4f668cb793b2", "event_id": "evt_8d32783c978fc24d", "occurred_at": "2026-10-06T13:05:02Z", "published_at": "2026-10-06T13:05:02Z", "first_seen_at": "2026-10-06T15:31:14.145306Z", "time_precision": "published", "update_type": "confirmation", "material_change": true, "title_zh": "Reflection终于交卷：501B模型Beam亮相，仍落后最新中国头部开源模型", "title_en": "Reflection终于交卷：501B模型Beam亮相，仍落后最新中国头部开源模型", "what_changed_zh": "新报道补充了 Reflection 自行公布的基准测试结果，显示 Beam 在 DeepSWE 上得分 44.4，接近 GLM-5.2（44.0），但落后于 Qwen 3.8-Max、GLM-5.3、Kimi K3 和 DeepSeek V4.1 Flash。训练方面，预训练使用 6144 块 GB300，不到 4 周处理 23.8 万亿 token；随后用 1.05 万块 GB300 进行 4 周强化学习，生成超过 1 亿条训练轨迹。Beam 仍在最后安全测试阶段，仅少量用户可提前使用；完整权重、技术报告和模型卡计划本月晚些时候以 Apache 2.0 协议发布。", "what_changed_en": "New coverage adds Reflection's own benchmark results showing Beam scoring 44.4 on DeepSWE, close to GLM-5.2's 44.0 but behind Qwen 3.8-Max, GLM-5.3, Kimi K3 and DeepSeek V4.1 Flash. Training details: pretraining on 6,144 GB300 GPUs processing 23.8 trillion tokens in under 4 weeks, followed by 4 weeks of reinforcement learning on 10,500 GB300 GPUs generating over 100 million rollouts. Beam is still in final safety testing with limited early access; full weights, technical report and model card are planned for later this month under Apache 2.0.", "current_state_zh": "Reflection AI 于 2026 年 10 月 5 日发布其首个前沿开源权重模型 Beam，这是一个总参数 5010 亿、激活参数 230 亿的稀疏 MoE 模型，面向编程、推理和智能体任务。新报道显示，Reflection 自行公布的基准测试中 Beam 在 DeepSWE 上得分 44.4，接近 GLM-5.2 但落后于最新的中国头部开源模型。Beam 仍在最后安全测试阶段，仅少量用户可提前使用；完整权重、技术报告和模型卡计划 2026 年 10 月晚些时候以 Apache 2.0 协议发布。训练使用 6144 块 GB300 预训练不到 4 周处理 23.8 万亿 token，再用 1.05 万块 GB300 进行 4 周强化学习，生成超过 1 亿条训练轨迹。", "current_state_en": "Reflection AI unveiled Beam on October 5, 2026, its first frontier open-weight sparse MoE model with 501B total and 23B active parameters for coding, reasoning and agentic tasks. New coverage shows Reflection's own benchmarks place Beam at 44.4 on DeepSWE, close to GLM-5.2 but behind the newest Chinese open-weight models. Beam is still in final safety testing with limited early access; full weights, technical report and model card are planned for later in October 2026 under Apache 2.0. Training used 6,144 GB300 GPUs for pretraining under 4 weeks on 23.8 trillion tokens, followed by 4 weeks of reinforcement learning on 10,500 GB300 GPUs generating over 100 million rollouts.", "detailed_summary_zh": "Nvidia-backed Reflection announced Beam, a 501B-parameter MoE open-weights model (23B active) for code, reasoning and agent tasks trained on thousands of GB300 GPUs, though its own benchmarks place it behind the newest Chinese open-source models and full weights are only due later this month under Apache 2.0.", "detailed_summary_en": "Nvidia-backed Reflection announced Beam, a 501B-parameter MoE open-weights model (23B active) for code, reasoning and agent tasks trained on thousands of GB300 GPUs, though its own benchmarks place it behind the newest Chinese open-source models and full weights are only due later this month under Apache 2.0.", "background_zh": "", "background_en": "", "community_discussion_zh": "", "community_discussion_en": "", "market_impact_zh": "", "market_impact_en": "", "importance_score": 7.0, "references": [], "confidence": 0.95, "story_ids": [ "telegram:theblockbeats:199137" ], "sources": [ { "url": "https://m.theblockbeats.info/flash/370455?from=telegram", "label": "theblockbeats", "source_type": "telegram", "official": false } ] }, { "update_id": "upd_91c9ea34f3d4142e", "event_id": "evt_8d32783c978fc24d", "occurred_at": "2026-10-16T14:58:02Z", "published_at": "2026-10-16T14:58:02Z", "first_seen_at": "2026-10-16T16:51:55.152680Z", "time_precision": "published", "update_type": "confirmation", "material_change": true, "title_zh": "英伟达传洽购的Reflection开源Beam权重：美国开源第二，落后Meta 9分", "title_en": "英伟达传洽购的Reflection开源Beam权重：美国开源第二，落后Meta 9分", "what_changed_zh": "Reflection 已将 Beam 1.0 的权重上传至 Hugging Face，兑现 10 月 5 日预告，采用 Apache 2.0 许可，同时开放 BF16、FP8、NVFP4 版本，支持 100 万 token 上下文。此外还开源了 Mirror CLI 编程智能体框架和 DFlash2 推测解码加速模型，并与 vLLM 团队合作优化推理，单请求解码每步耗时从 8.95ms 降至 3.30ms（2.71 倍提速）。Artificial Analysis 独立评测给 Beam 30 分，在美国开放权重模型中排名第二，但落后于中国模型。发布获 AWS、Google Cloud、英伟达等 40 多家合作伙伴支持。此前报道的英伟达进一步注资或收购 Reflection 的谈判仍在进行中。", "what_changed_en": "Reflection has uploaded the weights of Beam 1.0 to Hugging Face, fulfilling its October 5 preview, under Apache 2.0 with BF16, FP8 and NVFP4 versions and 1M-token context. It also open-sourced the Mirror CLI coding agent framework and DFlash2 speculative decoding accelerator, and worked with vLLM to optimize inference, cutting per-step decoding time for a single request from 8.95 ms to 3.30 ms (2.71x speedup). Artificial Analysis gives Beam a score of 30, ranking second among US open-weight models but behind Chinese models. The release has support from over 40 partners including AWS, Google Cloud and Nvidia. Nvidia's reported talks to further invest in or acquire Reflection are ongoing.", "current_state_zh": "Beam 1.0 的完整权重已于 2026 年 10 月 16 日前后在 Hugging Face 上以 Apache 2.0 协议发布，提供 BF16、FP8 和 NVFP4 版本，支持 100 万 token 上下文。Reflection 同时开源了 Mirror CLI 编程智能体框架和 DFlash2 推测解码加速模型，并与 vLLM 合作优化推理性能。Artificial Analysis 独立评测给 Beam 打 30 分，在美国开放权重模型中排名第二，但落后于 DeepSeek、GLM 等中国模型。发布获 40 多家合作伙伴支持。英伟达进一步注资或收购 Reflection 的谈判仍在进行中。", "current_state_en": "Beam 1.0's full weights were released on Hugging Face under Apache 2.0 around October 16, 2026, with BF16, FP8 and NVFP4 versions and 1M-token context. Reflection also open-sourced the Mirror CLI coding agent framework and DFlash2 speculative decoding accelerator, and worked with vLLM to optimize inference. Independent Artificial Analysis evaluation gives Beam a score of 30, ranking second among US open-weight models but behind Chinese models such as DeepSeek and GLM. The release has support from over 40 partners. Nvidia's reported talks to further invest in or acquire Reflection are ongoing.", "detailed_summary_zh": "Reflection released the open weights of its first model Beam 1.0 — a 501B-parameter MoE with ~23B active parameters and 1M-token context under Apache 2.0 — alongside the Mirror CLI agent framework, the DFlash2 speculative decoding accelerator, and vLLM inference optimizations that cut per-step decoding time by 2.71x.", "detailed_summary_en": "Reflection released the open weights of its first model Beam 1.0 — a 501B-parameter MoE with ~23B active parameters and 1M-token context under Apache 2.0 — alongside the Mirror CLI agent framework, the DFlash2 speculative decoding accelerator, and vLLM inference optimizations that cut per-step decoding time by 2.71x.", "background_zh": "", "background_en": "", "community_discussion_zh": "", "community_discussion_en": "", "market_impact_zh": "", "market_impact_en": "", "importance_score": 7.5, "references": [], "confidence": 0.95, "story_ids": [ "telegram:theblockbeats:200942" ], "sources": [ { "url": "https://m.theblockbeats.info/flash/372281?from=telegram", "label": "theblockbeats", "source_type": "telegram", "official": false } ] } ] }