AMAZINGINDEX.COM 日报快照
60.6
VOL. 2026.07
2026.07.16
← 返回 2026.07.16 日报
日报快照 · Daily Snapshot
NO. 009

Inkling开源975B MoE模型,支持百万上下文

#ARTICLE HackerNews 2026.07.16
推荐指数 72.0 NO. 009 · 2026.07.16
发布2026/07/15Score70Comments27

AI公司Scaled Foundations发布自研MoE架构大模型Inkling,975B总参数/41B激活参数,预训练于45万亿多模态token,支持1M上下文窗口,同步放出12B激活参数的轻量版Inkling-Small。这是少数完全开源权重、允许自由定制的超大规模模型,对需要长上下文和多模态能力的垂直场景开发者有直接价值。

Inkling的发布时间点卡在Llama 4争议之后、DeepSeek之后开源社区对'真开源'极度敏感的窗口期。1M上下文+全权重开放是差异化卖点,但团队背景(Pieter Abbeel的机器人/强化学习背景)暗示这模型可能不只是聊天工具,而是为具身智能和长期任务规划设计的基座。

当前开源MoE赛道Mistral和DeepSeek已经占据生态位,Inkling如果没有配套的推理优化工具链,实际落地会受限于激活参数仍达41B的算力门槛。建议关注Inkling-Small的评测表现,12B激活参数如果能在长文本任务上接近Llama 3 70B水平,可能是RAG和文档分析场景的黑马选项。

另外值得注意:声明里强调'extends human will and judgment'而非替代,这个措辞和Anthropic的constitutional AI形成微妙对照,后续看是否会在RLHF阶段引入人机协作的新范式。

意见分歧 29 条评论

核心争论:开源MoE模型能否在性能落后GLM 5.2的情况下,靠长上下文和多模态差异化胜出

ls_stats

America needs its own DeepSeek or Z.ai, a lot of people (myself included) root for open chinese models to win because they have no other choice. Thinking Machines might be it.

verdverm

Its not as good as GLM 5.2 for agentic workflows while also being bigger. Competition is going to be ruthless because the super low cost to switching. There is also AllenAi in the US, but they have yet to produce a model at this scale. Thankfully, new contenders can come out of nowhere and do well,

gkapur

It could be but there are a host of companies going after open weights models: Arcee, Reflection, Llama (TBD on Meta's focus on closed-source versus open-source), etc. That said, the fine-tuning API + open weight model at least is a semblance of a viable business that could work so I will be curious

替代方案: GLM 5.2DeepSeekZ.aiLlamaArceeReflectionAllenAiGemini 3.5 FlashDwarfStar
查看原文 →