Mira Murati's Inkling Hits the Open Source Scene

PromptCube3.com Novice 10h ago 81 views 11 likes 1 min read

Mira Murati's Inkling Hits the Open Source Scene
via aibase.com
Key points
  • Thinking Machines Lab released Inkling, a 975B MoE multimodal model with a 1M token context window.

  • The model shows exceptional math performance (97.1% on AIME2026) and strong native audio/visual capabilities.

  • Weights are now available on Hugging Face and via the Thinker API.
  • It's official: Mira Murati is back in the game with a powerhouse team largely poached from OpenAI. Inkling isn't just another model; it's a direct shot at the current open-source leaderboard. While it doesn't quite touch the absolute ceiling of closed-source giants in coding or complex reasoning, its multimodal integration is impressive.

    What stands out to me is the native audio and vision processing. Hitting 77.2% on MMAU puts it in the same league as Gemini 3.1 Pro, which is a huge win for the open-source community. The MoE architecture allows it to maintain a massive knowledge base while keeping active parameters low enough to be practical.

    Seeing this level of performance coming from a new lab so quickly suggests that the "OpenAI playbook" is being successfully replicated and open-sourced. Inkling provides a formidable base for developers who need high-end multimodal capabilities without being locked into a proprietary ecosystem.

    DeepSeekOpenAIAnthropicHugging Face

    All Replies (0)

    No replies yet — be the first!

    Write a Reply

    Markdown supported