第一项始终是图钉的来源。总体可信度是各份资料对上方所用开始和结束时间支持程度的加权平均;资料每比最新的一份旧 180 天,权重减半。
Thinking Machines[3] 的文章《Inkling: Our Open-Weights Model》发布于 2026 年 7 月 15 日,其中写道“Today we are advancing our mission by releasing a model we trained from scratch”;The Next Web[5] 于 7 月 15 日刊发报道,OpenRouter[4] 自 7 月 17 日起上架。
Model card: 975B total and 41B active parameters, 66-layer sparse MoE with 6 of 256 routed experts plus 2 shared, Apache 2.0, BF16 and NVFP4 weights, text/image/audio input with text output, and the benchmark table against Gemini 3.1 Pro, Claude Fable 5 and GPT 5.6 Sol.[1]
The lab's follow-up two weeks later: Inkling-Small, 276B total and 12B active, and the output pricing of $4.05 per 1 million tokens for Inkling against $1.20 for Inkling-Small.[1]
Router listing: 41B active of 975B, 524K context, $0.95 per 1 million input tokens and $4.05 per 1 million output tokens, released Jul 17, 2026.[1]
Independent launch-day report: the lab's first model, open-weight and not the strongest, distillation from other open models including Kimi K2.5, built in about nine months, and a Bridgewater fine-tuning result that is the partners' own evaluation.[1]
有遗漏或错误吗?用你自己的话说明:能佐证此图钉的链接、不同的开始或结束日期及理由,或缺失、有误的信息。AI 会对照此图钉的来源进行核实,搜索更好的来源,并添加任何支持你说法的页面。图钉自身的来源仍然最重要。AI 也会查看图片:显示的是别的东西或显示效果差的图片会被移到后面或替换。
原标题: Inkling Released