← 回總覽

Thinking Machines 开源多模态模型 Inkling,性能对比全面

📅 2026-07-16 10:24 meng shao 人工智能 4 分鐘 4281 字 評分: 86
Inkling Thinking Machines 开源模型 多模态 MoE
📌 一句话摘要 OpenAI 前 CTO Mira Murati 创立的 Thinking Machines 发布 975B 参数的 MoE 开源模型 Inkling,作者给出了其与主流开源及闭源模型的详细对比。 📝 详细摘要 Thinking Machines 正式开源了 Inkling 多模态模型(975B A41B MoE)。根据作者整理的基准测试对比:Inkling 强于 Nemotron 3 Ultra、Kimi K2.5,与 Kimi K2.6/DeepSeek V4 Pro 互有胜负,整体略弱于 GLM 5.2;在指令遵循上反超部分闭源模型,但在 agentic coding
Skip to main contentAudio 2 ![Image 1: LogoBest Blogs](https://www.bestblogs.dev/ "BestBlogs.dev")

Search Ctrl+K

Change language Switch ThemeSign In

Curated Daily BriefWeekly PicksTopicsWorld Cup Special SettingsHelp CenterCollapse

Narrow Mode

Thinking Machines Open-Sources Multimodal Model Inkling, Comprehensive Performance Comparison

Thinking Machines Open-Sources Multimodal Model Inkling, Comprehensive Performance Comparison

![Image 2: meng shao](https://www.bestblogs.dev/en/tweets?sourceId=SOURCE_65e681) meng shao

@shao__meng

美国开源模型 +1

来自 Thinking Machines (OpenAI 前 CTO @miramurati 的团队) 开源发布多模态模型 Inkling,975B/A41B MoE 架构 huggingface.co/thinkingmachin…

在开源模型中的位置

· 强于 Nemotron 3 Ultra、Kimi K2.5

· 与 Kimi K2.6 / DeepSeek V4 Pro 互有胜负:Inkling 在 IFBench、FORTRESS、SWEBench Verified 略胜;在 HLE、SimpleQA、Terminal Bench 略输

· 整体略弱于 GLM 5.2:GLM 5.2 在 HLE、AIME、SWEBench Pro、Terminal Bench、MCP Atlas、Tau 3 都更强,是当前开源组的"事实王者"

与闭源模型的差距

· 在 agentic coding / 综合推理 / 事实性 上,Claude Fable 5 和 GPT 5.6 Sol 明显领先一代(SWEBench Verified 95% vs 77.6%,HLE-with-tools 64.5% vs 46%,SimpleQA 71.6% vs 43.9%)

· 在 指令遵循 + 不过度拒绝 上,Inkling 反超部分闭源(IFBench 79.8% > Claude 63.5%;FORTRESS Benign 95.9% >> Claude 55.1%)Show More

!Image 3: Tweet image

!Image 4: Thinking Machines

#### Thinking Machines

@thinkymachines · 10h ago

Today, we are introducing Inkling.

Inkling reasons efficiently across text, image, and audio modalities. We are making the full weights available. thinkingmachines.ai/news/introduci…

Available today for fine-tuning on Tinker. Play with it in the Inkling Playground. 🧵Show More

418

1,377

10.4K

3.9M

Jul 16, 2026, 2:24 AM View on X

1 Replies

0 Retweets

0 Likes

530 Views ![Image 5: meng shao](https://www.bestblogs.dev/en/tweets?sourceid=65e681) meng shao @shao__meng

Follow

One Sentence Summary

Thinking Machines, founded by former OpenAI CTO Mira Murati, releases the open-source MoE model Inkling with 975B parameters. The author provides a detailed comparison with mainstream open-source and closed-source models.

Summary

Thinking Machines has officially open-sourced the Inkling multimodal model (975B A41B MoE). According to the author's benchmark comparison: Inkling outperforms Nemotron 3 Ultra and Kimi K2.5, trades wins and losses with Kimi K2.6 / DeepSeek V4 Pro, and is overall slightly weaker than GLM 5.2; it surpasses some closed-source models in instruction following but still lags significantly behind Claude Fable 5 and GPT 5.6 Sol in agentic coding and factuality.

AI Screening

86

Influence Score 1

Published Today

Language

Chinese

Tags

Inkling

Thinking Machines

Open-Source Model

Multimodal

MoE

Make your daily reading actually fit you.A daily brief built from the sources you follow. Get started free HomeDiscoverWorld CupSettings

查看原文 → 發佈: 2026-07-16 10:24:07 收錄: 2026-07-16 12:00:53

🤖 問 AI

針對這篇文章提問,AI 會根據文章內容回答。按 Ctrl+Enter 送出。