Event ·Multi-source

thinkymachines开源Inkling：975B总参/41B激活MoE，1M上下文多模态

Inkling是thinkymachines发布的首个开源MoE模型，总参数量975B，激活参数41B。它支持最高1M上下文长度，原生处理文本、图像和音频。该模型已获得SGLang和Miles的Day 0支持，采用短卷积、相对位置嵌入注意力和共享专家sink MoE等新架构。推理优化包括DFlash投机解码和MXFP8 KV缓存，并支持全参数和LoRA RL训练。

Current read

thinkymachines开源了一个超大MoE模型Inkling，总参975B但只激活41B，还支持1M上下文和图文音频理解，已经可以在SGLang上跑了，想玩大模型的朋友可以试试。

11 sources86° AI heatLast updated 2026/7/16 13:16:56

Evidence

11 sources

Primary sourceLMSYS Org (SGLang)

2026/7/16 13:16:56View source

Related source 1vLLM

2026/7/15 19:02:05View source

Related source 2Fireworks AI

2026/7/15 22:08:45View source

Related source 3Latent Space (swyx)

2026/7/16 06:18:05View source

Related source 4IT之家

2026/7/16 06:49:12View source

Related source 5elvis

2026/7/15 19:09:27View source

Related source 6The Rundown AI

2026/7/15 19:16:04View source

Related source 7marktechpost

2026/7/15 23:48:58View source

Related source 8shao__meng

2026/7/16 02:24:07View source

Related source 9歸藏(guizang.ai)

2026/7/16 03:57:41View source

Related source 10Simon Willison’s Weblog

2026/7/16 15:35:25View source

Conflict check

Dedup data only shows that these sources discuss the same event, not that they agree. There is no structured supporting or refuting evidence yet, so no conflict is inferred automatically.

Open the story