@Gorden_Sun
Thinking Machines发布Inkling：Mira Murati团队的首个开源多模态大模型 975B总参数、激活41B的MoE架构，Apache 2.0协议，支持文本、图像、音频输入，输出文本。
跑分处于开源第一梯队中上游，HLE带工具46.0%、SWEBench Verified 77.6%、GPQA 87.2%，与Kimi K2.6、DeepSeek V4 Pro同档，距离闭源模型还有差距。
这是首个1T参数级别的开源多模态模型，且支持在自家Tinker平台做微调定制。
模型：huggingface.co/thinkingmachin…
1 Replies
0 Retweets
1 Likes
753 Views 
One Sentence Summary
Thinking Machines has released Inkling, the first open-source multimodal large language model from Mira Murati's team, with 975B total parameters and 41B active parameters, supporting text, image, and audio input, and providing a HuggingFace download link.
Summary
Thinking Machines announced that its team led by Mira Murati has launched Inkling, the first open-source multimodal large language model. The model has a total of 975B parameters, with 41B active parameters using MoE architecture, and is licensed under Apache 2.0. It supports text, image, and audio input, and outputs text. On benchmarks, it scores 46.0% on HLE with tools, 77.6% on SWEBench Verified, and 87.2% on GPQA, placing it in the same tier as Kimi K2.6 and DeepSeek V4 Pro, but still behind top closed-source models. The model is said to be the first open-source multimodal model at the 1T parameter scale, and can be fine-tuned on the Tinker platform. The model has been released on HuggingFace at https://huggingface.co/thinkingmachines/Inkling.
AI Screening
81
Influence Score 2
Published Today
Language
Chinese
Tags
Thinking Machines
Inkling
Open-Source Multimodal Large Language Model
Mira Murati
MoE
Make your daily reading actually fit you.A daily brief built from the sources you follow. Get started free