Mira Murati’s Thinking Machines Lab launches Inkling to challenge Chinese AI
Thinking Machines Lab has released Inkling, a 975-billion parameter open-weight multimodal AI model designed for enterprise applications including video and audio analysis. The startup, founded by former OpenAI CTO Mira Murati, aims to provide a US-based alternative to international models, allowing for self-hosted deployment and domain-specific fine-tuning on custom infrastructure.
Key Takeaways
- 975-billion parameter MoE architecture with 41 billion active parameters and a 1-million-token context window
- Pretrained on 45 trillion tokens spanning text, image, audio, and video for native multimodal reasoning
- Deployment requires significant infrastructure, needing at least 2 TB of VRAM for the full BF16 checkpoint
- Features a 'reasoning-effort' setting (0.2 to 0.99) to balance processing depth against token generation speed
- Available through APIs on Together AI, Fireworks, and Databricks, with local support for vLLM and llama.cpp
Why It Matters
Inkling provides critical strategic diversification for enterprises that require self-hosted, multimodal AI but face regulatory or procurement barriers when using leading Chinese open-weight models like DeepSeek or GLM. By bridging the gap between proprietary frontier performance and the flexibility of open weights, Thinking Machines targets the high-volume 'intelligence ownership' market where domain-specific fine-tuning is paramount. For the streaming and video tech stack, the model's native multimodal training offers a localized path for building private, high-accuracy analysis agents without routing sensitive media through closed third-party APIs. Watch for the release of Inkling-Small as a more economical benchmark for production-scale deployments.
Additional Context
The launch of Inkling arrives as the open-weight AI sector transitions into a primary competition between U.S. and Chinese developers. Per SCMP and Vercel (July 2026), open-weight models recently surged to 29% of all production AI tokens, with Chinese models like Zhipu’s GLM-5.2 and DeepSeek V4 Flash capturing over 20% of gateway traffic. These models often operate at roughly one-tenth the cost of premium closed alternatives from OpenAI or Anthropic. This shift has placed immense pressure on Western labs to provide comparable open weights that maintain domestic compliance and data residency for global banks and healthcare providers. Thinking Machines has faced a high-stakes path to this release following its February 2025 founding. Per Reuters and Bloomberg (June 2026), the startup secured a record $2 billion seed round at a $12 billion valuation in July 2025, backed by Nvidia and Andreessen Horowitz. However, the company navigated significant internal volatility in early 2026, including the departure of three original co-founders and key researchers who returned to OpenAI or Meta. Despite these talent losses, Thinking Machines secured massive compute agreements with Nvidia and Google Cloud to support the training of Inkling on next-generation hardware like Nvidia's B300 and H200 GPUs. Market analysis from OpenRouter (June 2024–2026) highlights that while U.S. firms like Nvidia have released open-weight models such as Nemotron 3 Ultra, the mid-2026 leaderboard was largely dominated by Chinese equivalents in coding and reasoning tasks. The introduction of Inkling’s 'interaction model' design aims to differentiate through real-time collaboration, a strategy Murati has championed to move beyond traditional, turn-based prompting. This approach reflects a broader trend noted by TCV (July 2026) where the technical challenge has shifted from merely increasing model scale to optimizing inference and private data integration for reliable production use.
Read full article at infoworld.com
Get this in your inbox → Subscribe
Enjoy our coverage?
Add StreamingMeme as a preferred source on Google to see more of our streaming news at the top of your Search results.
Add as preferred source