Thinking Machines launched Inkling on July 15. The San Francisco startup founded by former OpenAI CTO Mira Murati released the full weights of its first AI model. This move signals a fresh push from Western labs to counter the flood of high-performing open-weight systems coming from China.
Inkling stands out as a Mixture-of-Experts transformer. It counts 975 billion total parameters but activates just 41 billion for any given task. The model handles a context window of up to 1 million tokens. It draws from pretraining on 45 trillion tokens spanning text, images, audio and video. And a smaller variant called Inkling-Small, with 12 billion active parameters, appeared in preview too.
But size tells only part of the story. The company positioned Inkling as a broad, balanced foundation rather than a benchmark chaser. It reasons natively across modalities. It offers controllable thinking effort that lets users trade performance for lower latency and cost. One chart in the announcement showed Inkling matching certain scores from NVIDIA’s Nemotron 3 Ultra while using roughly one-third the tokens on agentic coding tasks. Efficiency like that matters when models run millions of times inside enterprise workflows.
Executives at Thinking Machines argue customization beats one-size-fits-all systems. They point to real-world applications where tailored models deliver better results at lower expense. Bridgewater Associates, for instance, used the startup’s Tinker platform to build a custom version of Alibaba’s Qwen model. That version outperformed top proprietary systems while slashing costs dramatically, according to Investing.com.
The open-weight space has tilted heavily toward Chinese developers. Hugging Face reported in its Spring 2026 review that models from China accounted for 41 percent of downloads over the prior year. They surpassed U.S. offerings in both monthly and cumulative metrics. Stanford’s 2026 AI Index noted the performance gap between leading closed and open models had narrowed to 3.3 percent by March. Six of the top ten models on the Arena leaderboard remained closed, yet the floor for open systems kept rising. Forbes captured the shift in April when it observed open source AI moving from sideshow to strategy.
Meta’s decision to pull back from fully open releases after Llama 4 disappointed many in the community. That vacuum accelerated adoption of alternatives from Alibaba, DeepSeek and others. Businesses sought cheaper, customizable options without the governance headaches of closed frontier models. Yet concerns linger. Regulators and some executives worry about safety when weights circulate freely. Thinking Machines trained Inkling to an internal standard of safe behavior. It evaluated the system on benchmarks such as FORTRESS and StrongREJECT. External red teams conducted additional tests. The company said it would continue studying safety for customizable models.
So what makes this release different? Inkling avoids narrow optimization for any single benchmark family. Charts shared by the lab placed it competitively against GLM 5.2, Nemotron 3 Ultra, GPT-5.6 Sol and Claude Fable 5 across agentic, reasoning, coding, vision and audio evaluations. On Design Arena’s Agentic Web Dev leaderboard, blinded human raters ranked Inkling among the strongest open-weight entries. It built functional web apps in one shot. It produced polished nine-page PDFs with cohesive styling. It even iterated 40 times on a multiplayer snake game when paired with a reviewer model.
Multimodal performance drew particular attention. Inkling scored 73.5 percent on MMMU Pro standard, 78.1 percent on Charxiv RQ and 91.4 percent on VoiceBench at high effort. Those numbers sit near or above several competing open systems while trailing the very top closed models. The architecture interleaves sliding-window and global attention layers. It uses a sigmoid-based router with 256 experts. Such design choices support efficient inference and the controllable effort feature that developers can dial from 0.2 to 0.99.
Recent coverage highlights the strategic bet. TechCrunch reported how the model challenges the assumption that generalist systems must dominate. It noted Inkling’s emphasis on calibrated answers that flag uncertainty. The piece also referenced stalled talks of a $50 billion funding round and partnerships with NVIDIA. Meanwhile Wired detailed the company’s origins. Founded in February 2025 by Murati, John Schulman and Lilian Weng, Thinking Machines secured a seed round that valued it at $12 billion. The lab released Tinker last October to let users customize models. It has since previewed interaction models for natural voice and vision collaboration.
One internal experiment stood out. Inkling fine-tuned itself on Tinker. The model wrote its own training job, executed it and evaluated the outcome. During earlier training it had stripped explanatory grammar from chain-of-thought reasoning because the tokens counted as overhead. Engineers reinstated natural language steps to preserve explainability. “It determined that the grammar was overhead, which is interesting,” a company source told Wired. They added that the team “reinstated natural language reasoning to make the models’ decisions more explainable.”
Analysts see broader momentum. OpenRouter’s June 2026 review listed DeepSeek V4 Flash, GLM 5.2, MiniMax M3 and NVIDIA’s Nemotron 3 Ultra as the open-weight models that mattered most at the time. Inkling enters that conversation as a native Western alternative with strong multimodal roots. Nathan Lambert wrote in April on Interconnects.ai that U.S. labs could regain ground in adoption metrics starting in 2027 thanks to releases like Google’s Gemma 4 and Arcee AI’s focus. Chinese labs, he predicted, might face funding pressure first.
Yet companies still pay for closed models. A Medium post published three days ago asked why enterprises reach for credit cards when open LLMs have caught up in 2026. The author pointed to reliability, support and integration ease as persistent advantages for proprietary offerings. Open-weight systems shine when teams need control, fine-tuning or data privacy. They falter when users demand the absolute highest ceiling without extra engineering.
Thinking Machines employs about 200 people. It cultivates a culture of continuity rather than star power. Murati has kept a lower profile than during her OpenAI days. The lab’s mission statement speaks of building AI that extends human will and judgment. Releasing weights fits that vision. It lets developers and enterprises adapt the system to their data and workflows instead of depending on a handful of labs in San Francisco or Beijing.
Availability matters too. Inkling works today for fine-tuning inside Tinker. A dedicated playground lets users chat with it directly. Support spans 64K and 256K context lengths initially. The smaller Inkling-Small preview hints at future releases that could run on more modest hardware. Full weights for the small model will follow.
Industry reaction on X poured in quickly. Users called it one of the more significant Western open releases of the year. Some praised its competitive standing against GPT-5.6 Sol on design benchmarks. Others noted the Apache 2.0 license and native multimodal training as big draws. One post highlighted how the model was pretrained from scratch rather than distilled heavily from closed systems.
Challenges remain. The performance gap, though narrow, still favors certain closed models on the hardest reasoning tasks. Safety evaluations for open weights draw scrutiny as capabilities advance. And the economics of training at this scale require deep pockets. Thinking Machines raised $2 billion in 2025 at a $10 billion valuation with backing from NVIDIA, AMD, ServiceNow, Cisco and Andreessen Horowitz. Those resources helped.
Looking ahead, the lab plans to expand the model family. It will study safety implications of highly customizable systems. It aims to make fine-tuning accessible for more use cases. If Inkling gains traction among developers, it could slow the shift toward Chinese open models and give Western enterprises a trusted alternative.
The release arrives at a moment when open-weight AI has moved firmly into mainstream strategy. Enterprises experiment with self-hosted systems to cut costs and retain control. Researchers fine-tune them for specialized domains. Regulators debate export controls and safety standards. Against that backdrop, a well-rounded, multimodal model from a credible new lab carries weight. It doesn’t claim to lead every leaderboard. It offers something rarer: a flexible starting point that teams can shape to their own needs.
Whether that proves enough to shift market share will unfold over the coming months. Benchmarks will be run. Models will be forked and improved. Enterprises will test it inside their workflows. For now, Inkling stands as a notable entry. It reminds the industry that specialization, efficiency and openness still hold appeal even as the biggest labs chase ever-larger generalist systems.
WebProNews is an iEntry Publication