Inkling: Our Open-Weights Model
thinkingmachines.ai
July 15, 2026
20 min read
🔥🔥🔥🔥🔥
77/100
Summary
Inkling is a Mixture-of-Experts transformer model with 975 billion total parameters and 41 billion active parameters. The model is released with full weights available for customization, supporting interactive collaboration and user-defined modifications.
Key Takeaways
- Inkling is a Mixture-of-Experts transformer model with 975 billion total parameters and 41 billion active parameters, pretrained on 45 trillion tokens across various media types.
- Inkling supports a context window of up to 1 million tokens and is designed for multimodal capabilities, allowing reasoning over text, images, and audio.
- Inkling is available for fine-tuning on the Tinker platform, with a developer interface called Inkling Playground for customization and interaction.
- The model is trained to perform well across diverse tasks, including coding and agentic tool use, making it suitable for a wide range of real-world applications.
Coverage & discussions
Hacker News · thinkingmachines.ai
Community discussions
Community Sentiment
Positives
- Finally, an American open weights model that competes with Chinese counterparts — this could shift the landscape for accessible AI solutions.
- The multi-modal capabilities, especially for audio, are a game changer. This could unlock tons of applications that are currently limited by tech.
- Fine-tuning APIs are straightforward, hinting at a potential goldmine for developers looking to customize AI without the usual headaches.
- Some commenters are thrilled about the long context capabilities, which could enhance performance in complex tasks.
- This model is being touted as the best American open weights option, setting a new benchmark for others to aspire to.
Concerns
- Skeptics point out that benchmark comparisons seem to favor this model, with some saying it performs worse than established counterparts like KimiK2.7.
- There's concern about whether the model's capabilities are truly competitive or if it's just hype — a few users are still leaning on older models.
- Some are wary about the actual performance drop-off after certain context lengths, questioning the reliability for extensive tasks.
- Critics are not convinced that the synergy between the open weights model and fine-tuning API is fully realized, leaving doubt about its viability.
