Thinking Machines Lab releases Inkling, a 975B-parameter open-weights multimodal model under Apache 2.0 license
Open source release ● Confirmed 92% confidence first seen
Thinking Machines Lab released Inkling, an open-weights Mixture-of-Experts multimodal model with 975 billion total parameters (41 billion active) trained on 45 trillion tokens, supporting a 1 million token context window and available under Apache 2.0 license on Hugging Face. The release includes a smaller 276-billion-parameter variant and is positioned as a strong base model for fine-tuning across diverse applications. The model represents one of the strongest U.S.-based open-weights models released to date, designed for developer customization and deployment with controllable computational efficiency.
Decision brief
- What changed
- Thinking Machines Lab (founded by Mira Murati) released Inkling, an open-weights Mixture-of-Experts multimodal model with 975 billion total parameters (41 billion active), trained on 45 trillion tokens, supporting a 1 million token context window, under an Apache 2.0 license on Hugging Face and its Tinker platform. A smaller 276-billion-parameter variant is also available.
- Why it matters
- This gives enterprises a permissively licensed, U.S.-based open-weights model strong enough to serve as a fine-tuning base, reducing dependence on proprietary APIs or Chinese open-weights alternatives. Because it's positioned as a base model rather than a frontier system, decision-makers should weigh it for customization and cost-controlled deployment rather than as a drop-in replacement for top-tier commercial models.
- Evidence
- Three independent outlets (Latent Space/AINews, TLDR Dev, Simon Willison's blog) report consistent technical specifications—975B/41B parameters, 45T training tokens, 1M context, Apache 2.0 license—indicating reliable, cross-verified basic facts about the release.
- What remains uncertain
- No independent benchmark comparisons are provided beyond a general note that it trails top Chinese open models; actual inference costs, hardware requirements, fine-tuning support quality on the Tinker platform, and real-world performance across text/image/audio tasks remain unverified.
- Monitor next
- Watch for independent benchmark results and early enterprise fine-tuning case studies that clarify Inkling's real-world performance and total cost of deployment relative to competitors like Nemotron and Gemma.
Analytical support, not advice — assumptions and open questions stated above.