Thinking Machines open sources first multimodal language model, Inkling, focused on low cost and 'resistance to censorship'

2026-07-15 · Source: VentureBeat · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Software Development & Engineering · Depth: Advanced, medium

Summary

Thinking Machines, an American AI startup founded by former OpenAI CTO Mira Murati, released Inkling on July 15, 2026, its first open-source multimodal language model. Licensed under Apache 2.0, Inkling features 975 billion total parameters (41 billion active) and natively processes text, images, and audio. It boasts high, if sub state-of-the-art, performance for open-weights models, scoring 77.6% on SWE-bench Verified and 91.4% on VoiceBench, though it trails some Chinese rivals and closed-source models in peak reasoning. A key differentiator is its "controllable thinking effort" mechanism, allowing developers to adjust reasoning budget from 0.2 to 0.99 to balance cost and performance. Inkling also demonstrates "censorship non-compliance" on sensitive topics while maintaining 98.6% safety on malicious queries. A lighter 276-billion-parameter version, Inkling-Small, is also available.

Key takeaway

For AI Architects evaluating open-source models for agentic workloads, Inkling presents a compelling option due to its Apache 2.0 license and native multimodality. You can programmatically control cost-performance via its "controllable thinking effort" and benefit from its resistance to censorship for sensitive applications. However, be prepared to integrate external moderation tools for specific safety policies, as internal evaluations noted vulnerabilities to role-play prompts.

Key insights

Inkling offers a natively multimodal, open-source foundation model with programmatic cost-performance control and resistance to censorship.

Principles

• Balance cost and performance through adjustable reasoning.
• Design for censorship resistance without compromising safety.
• Natively fuse modalities for true multimodal reasoning.

Method

Inkling employs a Mixture-of-Experts (MoE) architecture with encoder-free early fusion for text, audio (dMel spectrograms), and visual (40x40 pixel patches) data, using relative positional embeddings and a "controllable thinking effort" mechanism.

In practice

• Adjust "thinking effort" (0.2-0.99) for cost/performance.
• Deploy Inkling-Small for low-latency, cost-sensitive tasks.
• Integrate external moderation tools like Llama Guard.

Topics

• Inkling Model
• Multimodal AI
• Open-Source LLMs
• Censorship Resistance
• MoE Architecture
• Cost-Performance Control

Best for: CTO, VP of Engineering/Data, Machine Learning Engineer, Director of AI/ML, AI Architect, AI Engineer

Related on AIssential

• Thinking Machines Lab’s First Model Gives Nod to Chinese Open Source AI — Thinking Machines Lab's Inkling model demonstrates leveraging established open-source architectures, specifically DeepSeek V3, for new AI development.
• Complete Guide to Thinking Machines Inkling — Inkling is an open-weights multimodal MoE model prioritizing customization and domain-specific fine-tuning over benchmark supremacy.
• Ex-OpenAI CTO Murati's Thinking Machines drops Inkling, a 975B parameter model that leads US labs but trails China — Inkling leads U.S. open-weights models in agentic tasks and efficiency but struggles with factual accuracy and cost.
• Mira Murati’s Thinking Machines drops Inkling, an open-weights model anyone can access — Open-weight multimodal models like Inkling offer customizable, cost-effective alternatives to proprietary AI systems.
• Introducing Gemma 4 12B: a unified, encoder-free multimodal model — Gemma 4 12B offers efficient, encoder-free multimodal AI, enabling advanced agentic reasoning on consumer laptops.
• Thinking Machines Lab Drops Its First Model — Thinking Machines Lab launched Inkling, a large open-source multimodal model, challenging established AI players with a vision for decentralized AI.

See Counsel's argued verdicts on the open AI decisions leaders are weighing →

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by VentureBeat.