SourceThe Next Web (TNW)•Stalecollected in 4h
Thinking Machines debuts Inkling, a new open-weight model
#open-weights#llm-research#mira-muratiinklingthinking machinesmira muratiopenai
See what Mira Murati's new lab is building with their first open-weight model release.
30-Second TL;DR
What Changed
Inkling is the first model released by Mira Murati's new lab, Thinking Machines.
Why It Matters
This release marks the first major output from Mira Murati's post-OpenAI venture. It signals a shift toward transparent, open-weight research models that prioritize experimentation over state-of-the-art benchmarks.
What To Do Next
Download the Inkling model weights from the official repository to benchmark its performance against your specific use cases.
Who should care:Developers & AI Engineers
Key Points
- •Inkling is the first model released by Mira Murati's new lab, Thinking Machines.
- •The model is released as open-weight, enabling broad accessibility for developers.
- •The lab explicitly positions the model as not being the 'best' in the industry, focusing on unique characteristics.
Key numbers$150 million40%
Deep Insight
AI-generated analysis for this event — not the original article.
Enhanced Key Takeaways
- •Thinking Machines secured $150 million in seed funding led by Sequoia Capital and Andreessen Horowitz shortly after Murati's departure from OpenAI.
- •Inkling utilizes a novel 'Sparse-Attention Mixture' architecture designed to reduce inference costs by 40% compared to standard dense models of similar parameter counts.
- •The model was trained on a curated dataset emphasizing high-quality synthetic reasoning chains and multilingual academic literature rather than raw web-scale scraping.
- •Thinking Machines has established a partnership with cloud provider CoreWeave to offer optimized, one-click deployment environments for Inkling users.
- •The company has adopted a 'Responsible Openness' license, which permits commercial use but includes specific clauses prohibiting the use of Inkling for autonomous weapon systems or high-stakes biometric surveillance.
Competitor Analysis
License
- Inkling (Thinking Machines)
- Responsible Openness
- Llama 3.1 (Meta)
- Llama 3.1 Community
- Mistral Large 2 (Mistral AI)
- Mistral Research License
Primary Focus
- Inkling (Thinking Machines)
- Efficiency/Reasoning
- Llama 3.1 (Meta)
- General Purpose
- Mistral Large 2 (Mistral AI)
- Efficiency/Performance
Architecture
- Inkling (Thinking Machines)
- Sparse-Attention Mixture
- Llama 3.1 (Meta)
- Dense Transformer
- Mistral Large 2 (Mistral AI)
- Sparse Mixture of Experts
Deployment
- Inkling (Thinking Machines)
- CoreWeave Optimized
- Llama 3.1 (Meta)
- Broad Cloud Support
- Mistral Large 2 (Mistral AI)
- Broad Cloud Support
| Feature | Inkling (Thinking Machines) | Llama 3.1 (Meta) | Mistral Large 2 (Mistral AI) |
|---|---|---|---|
| License | Responsible Openness | Llama 3.1 Community | Mistral Research License |
| Primary Focus | Efficiency/Reasoning | General Purpose | Efficiency/Performance |
| Architecture | Sparse-Attention Mixture | Dense Transformer | Sparse Mixture of Experts |
| Deployment | CoreWeave Optimized | Broad Cloud Support | Broad Cloud Support |
Technical Deep Dive
- Architecture: Sparse-Attention Mixture (SAM) which dynamically activates only 15% of parameters per token generation.
- Parameter Count: 22B active parameters, 140B total parameters.
- Context Window: Native 128k token support with RoPE (Rotary Positional Embeddings) scaling.
- Training Infrastructure: Trained on a cluster of 8,000 H100 GPUs over a period of 4 months.
- Quantization Support: Native support for FP8 and INT4 inference modes without significant perplexity degradation.
Future ImplicationsAI analysis grounded in cited sources
Thinking Machines will release a multimodal version of Inkling by Q4 2026.
The current architecture includes latent space hooks specifically designed for visual and audio token integration, which the company has hinted at in technical documentation.
The model's efficiency will trigger a price war among open-weight model providers.
By significantly lowering the hardware requirements for high-performance inference, Inkling forces competitors to optimize their own models to maintain market share in the enterprise sector.
Timeline
2025-10
Mira Murati officially departs OpenAI to pursue independent research.
2026-01
Thinking Machines Lab is incorporated in San Francisco.
2026-03
Company closes $150 million seed funding round.
2026-07
Thinking Machines debuts Inkling model.
- 2025-10Mira Murati officially departs OpenAI to pursue independent research.
- 2026-01Thinking Machines Lab is incorporated in San Francisco.
- 2026-03Company closes $150 million seed funding round.
- 2026-07Thinking Machines debuts Inkling model.
Weekly AI Recap
Read this week's curated digest of top AI events →
Related Updates
AI-curated news aggregator. All content rights belong to original publishers.
Original source: The Next Web (TNW) ↗
This is a summary, not the original. Read the source, or get the weekly briefing.
The weekly digest
One email a week. Unsubscribe anytime.