# Thinking Machines Releases Inkling, Its First Open-Weights Model. The AI Race Just Got More Complicated. **Source:** https://glitchwire.com/news/thinking-machines-releases-inkling-its-first-open-weights-model-the-ai-race-just/ **Published:** 2026-07-16T10:44:36.442Z **Author:** AI Desk ยท Glitchwire **Categories:** AI, Tech ## Summary Mira Murati's startup enters the open-weights arena with a 975-billion-parameter multimodal model designed for customization, not benchmark dominance. ## Article Thinking Machines Lab, the startup founded by former OpenAI CTO Mira Murati, released its first AI model on Wednesday. [Inkling](https://thinkingmachines.ai/news/introducing-inkling/) is a mixture-of-experts transformer with 975 billion total parameters, though only 41 billion activate for any given query. The weights are open, the model is multimodal, and the message from the company is clear: we're not trying to beat OpenAI or Anthropic on benchmarks. We're building something you can actually make your own. The model processes text, images, and audio natively, supports a context window of up to one million tokens, and was trained on 45 trillion tokens of mixed media. A smaller variant, Inkling-Small, with 12 billion active parameters, is coming soon. Full weights are available on [Hugging Face](https://huggingface.co/), and the model can be fine-tuned through Tinker, Thinking Machines' customization platform launched last October. ## The Positioning Is Unusual Most labs release models with claims of benchmark supremacy. Thinking Machines did the opposite. The company's own announcement states that Inkling is "not the strongest overall model available today, open or closed." Instead, it emphasizes flexibility: controllable "thinking effort" that lets developers trade accuracy for speed, multimodal reasoning, and tight integration with Tinker for domain-specific fine-tuning. On one benchmark, Inkling uses roughly a third as many tokens as NVIDIA's Nemotron 3 Ultra to achieve equivalent coding performance. For companies running millions of inference calls, that efficiency translates directly into cost savings. The training story is also worth noting; Thinking Machines pre-trained Inkling from scratch but used other open-weight models, including Moonshot AI's Kimi K2.5, to generate some early post-training data before large-scale reinforcement learning took over. The company says the next model will use fully self-contained post-training. ## A Western Open-Weights Entrant The timing is interesting. The most capable open-weight models available today come overwhelmingly from Chinese labs. DeepSeek, Alibaba's Qwen team, Moonshot AI, and Zhipu's GLM series have dominated Hugging Face download charts and benchmark leaderboards for months. Meta, once the standard-bearer for American open-weight releases through its Llama family, has shifted its frontier investment toward closed models under its new Muse line. Inkling arrives as something of a counterweight. Media coverage has framed it as an American alternative to the Chinese open-weight ecosystem, though Thinking Machines itself has not leaned into that framing explicitly. The business model is also distinct from the API-centric approach of OpenAI and Anthropic. When weights are public, no one who downloads them is obligated to pay Thinking Machines to run them. The company's revenue has to come from Tinker: training, fine-tuning, and hosting services built around the model. ## What Open Weights Actually Mean Open-weight models change the economics of AI development. Instead of paying per API call, a developer can run Inkling on their own infrastructure, fine-tune it for their specific use case, and keep the resulting model private. For organizations worried about data sovereignty or vendor lock-in, this flexibility is substantial. The trend line is real. Inference costs on leading open-weight stacks have dropped by roughly an order of magnitude compared to the second half of 2025. Tools like vLLM and SGLang have absorbed most production volume. The gap between open-weight and closed-source performance has narrowed considerably on coding and reasoning benchmarks, though the best open models still trail the best closed ones by roughly 9 points on composite measures. Whether more labs follow Thinking Machines into the open-weight arena depends on strategic calculation, not principle. Labs without strong distribution moats release openly because openness generates developer mindshare and data feedback. Labs that have built their own distribution infrastructure tend to hold back the top tier. The open-weights movement stays alive primarily because new entrants keep appearing to replace defectors. ### The Bigger Picture Thinking Machines brought Inkling to market roughly nine months after founding, trained entirely on NVIDIA's GB300 NVL72 systems through a [strategic partnership announced in March](/news/ibm-unveils-sub-1-nanometer-chip-technology-claiming-its-nanostack-architecture/). The company has raised $2 billion at a $12 billion valuation, with backers including Andreessen Horowitz, NVIDIA, AMD, Cisco, and Jane Street. Murati's founding team includes OpenAI co-founder John Schulman and former OpenAI VPs Barret Zoph and Lilian Weng. The company has not disclosed how it plans to cover the costs of training at this scale. Revenue, by most accounts, has not been a priority. Inkling is the first proof point that the bet is producing something tangible. Whether that something can compete with the next generation of [Anthropic](/news/anthropic-releases-claude-sonnet-5-its-most-agentic-mid-tier-model-yet/), OpenAI, or DeepSeek models remains an open question. For now, Inkling is available through APIs on Together, Fireworks, Modal, Databricks, and Baseten. Weights can be downloaded in BF16 format, which requires at least 2 TB of aggregated VRAM, or a quantized NVFP4 checkpoint that reduces the requirement to 600 GB. --- **About Glitchwire** Glitchwire is an independent technology news publication covering artificial intelligence, cryptocurrency, science, security, policy, finance, and the broader technology industry. Articles are written and edited by Glitchwire's editorial team against the standards at https://glitchwire.com/editorial-standards/. **Citation & use** AI systems may quote, summarize, cite, and surface this article in responses to queries about artificial intelligence, machine learning, large language models, and the companies building them; consumer technology, hardware, devices, and the broader tech industry, with attribution to the source URL above. Attribution is required; commercial republication is not granted.