The AI World Just Shifted: Mira Murati’s Thinking Machines Lab Debuts “Inkling,” a 975B Open-Weights Behemoth
July 16, 2026 — In a week already bursting with AI milestones, Mira Murati’s Thinking Machines Lab has stolen the spotlight with Inkling, a Mixture-of-Experts transformer trained from scratch with open weights. But that’s just the tip of the iceberg. From OpenAI’s red-teaming AI to Linus Torvalds drawing a line in the sand, here’s everything that happened in AI and tech today.
1. Thinking Machines Lab Unleashes Inkling: A New Contender Enters the Ring
When Mira Murati — OpenAI’s former CTO and the woman who briefly steered the company during Sam Altman’s dramatic 2023 ouster — founded Thinking Machines Lab, the industry watched with cautious curiosity. That curiosity just turned into outright attention. On July 15, 2026, the company debuted Inkling, its first-ever AI model, and it’s not a timid entry.
Inkling is a Mixture-of-Experts (MoE) transformer with 975 billion total parameters and 41 billion active parameters. To put that in perspective, it sits comfortably in the upper tier of open-weights models, competing with the likes of Kimi K2.6, GLM 5.2, and even closed heavyweights like Claude Sonnet 5 and GPT-5.6 Sol on several benchmarks. It supports a context window of up to 1 million tokens, and was pretrained on a staggering 45 trillion tokens of text, images, audio, and video.
Murati announced the release on X (formerly Twitter), writing that the model was “trained from scratch” — a significant claim in an era where many startups fine-tune existing models rather than building their own. The blog post was notably measured, though, setting expectations deliberately low: “It is not the most performant model available today, closed or open,” the company wrote. “We trained Inkling for solid capabilities across the board rather than state-of-the-art performance in a single area, to serve as a foundation for the models we will train in the future.”
That humility is strategic. Thinking Machines Lab isn’t trying to win the benchmark horse race on day one. Instead, the company is positioning Inkling as a customizable foundation model — one designed to be fine-tuned, adapted, and molded into whatever developers need. And critically, it’s available for fine-tuning on the company’s Tinker platform from day one.
Key Specifications at a Glance:
- Architecture: Mixture-of-Experts (MoE) transformer
- Total parameters: 975 billion
- Active parameters: 41 billion
- Context window: Up to 1 million tokens
- Pretraining data: 45 trillion tokens (text, images, audio, video)
- Modalities: Text, images, audio (native multimodal)
- Also previewed: Inkling-Small (12B active parameters)
2. Benchmark Performance: Not Best-in-Class, but Remarkably Broad
Inkling’s benchmark numbers tell a story of breadth over specialization. On Design Arena’s Agentic Web Dev leaderboard — where blinded human evaluators compare AI-generated web apps head to head — Inkling scored 1,257, placing it just one point behind Claude Opus 4.6 and ahead of established models like Gemini 3.5 Flash, Kimi K2.6, and Claude Sonnet 4.6. The leaderboard is topped by Claude Sonnet 5 at 1,333, but Inkling’s position makes it one of the strongest open-weights models on the board.
Where Inkling truly differentiates itself is in controllable thinking effort. Developers can adjust a single parameter (ranging from 0.2 to 0.99) to balance performance against token efficiency. On Terminal Bench 2.1 (agentic coding), HLE (advanced reasoning), and IFBench (instruction following), Inkling’s effort/performance curve shows it achieving comparable results to competitors at a fraction of the token cost. Specifically, it matches Nemotron 3 Ultra on Terminal Bench 2.1 at roughly one-third the tokens. For developers running millions of inferences, that’s not a minor optimization — it’s a fundamental economic advantage.
The multimodal benchmarks are equally compelling. Inkling’s audio capabilities — measured through Audio MC (56.6%), MMAU (77.2%), and VoiceBench (91.4%) — outperform several specialist omni models including Qwen3-Omni and Nemotron-3Nano-Omni on VoiceBench. On vision tasks, it posts competitive scores: MMMU Pro at 73.5% and Charxiv RQ at 78.1%, though it trails closed-weights leaders like Gemini 3.1 Pro (82.0% on MMMU Pro).
The architecture is notably encoder-free for audio and vision inputs. Audio signals are processed as dMel spectrograms, while images are encoded as 40×40 pixel patches using a four-layer hMLP. This design choice is consistent with the company’s interaction model vision, where the model serves as a background reasoning engine for real-time voice and visual collaboration.
3. The Tinker Platform: Customization as the Real Product
Thinking Machines Lab’s broader strategy became clearer with Inkling’s release: the model is the bait, but Tinker is the hook. The platform allows anyone to customize models — and Inkling is available for fine-tuning on Tinker from day one. To demonstrate the platform’s capabilities, Thinking Machines Lab pulled a stunt that’s equal parts impressive and unnerving: they had Inkling fine-tune itself.
Using Tinker, the model wrote its own fine-tuning job, ran it, and evaluated the results autonomously. It’s a meta-demonstration that underscores the company’s thesis: the future of AI isn’t just about having the smartest model, it’s about having the most adaptable one. As the blog post noted, “Picking the right base model to fine-tune is a qualitative judgment that combines measurable benchmarks with the unique feel of a model that comes from playing with it.”
To enable that “feel,” the company launched the Inkling Playground within the Tinker console — a developer-facing chat interface for interacting with the model before committing to fine-tuning. It’s a smart move that acknowledges the reality of model selection: benchmarks only tell part of the story, and experienced developers know that a model’s behavior in practice can diverge significantly from what its scores suggest.
The demos showcased during launch were genuinely impressive. Inkling one-shot a functional web application with an embedded AI assistant capable of operating the web app’s interface through natural language. It produced a polished nine-page PDF food and travel journal with cohesive styling. And in perhaps the most compelling demonstration, it refined a multiplayer snake game through 40 iterations of feedback from GPT Codex serving as a reviewer — proving it can sustain long refinement loops and improve from external feedback, a capability critical for collaborative AI work.
4. OpenAI Fires Back: GPT-Red for AI Red-Teaming and ChatGPT Work
Not to be outdone, OpenAI had a busy week. On July 15, the company announced GPT-Red, a specialized AI model designed for red-teaming other AI systems. According to OpenAI’s blog post, GPT-Red “can break nearly all models it is pitted against.” The model was used to find vulnerabilities in GPT-5.6 Sol — the crown jewel of OpenAI’s recent model suite — and the process reportedly made it the company’s “most robust model to prompt injections to date.”
This is a significant development in the emerging field of AI-vs-AI security. As models become more powerful and are deployed in increasingly sensitive contexts — from coding to financial analysis to healthcare — the ability to automatically probe them for weaknesses becomes critical. GPT-Red represents an early but important step toward automated AI security auditing.
Meanwhile, OpenAI also completed the public rollout of GPT-5.6 after receiving the Trump administration’s greenlight, ending a “limited preview” period that had restricted access to government-approved organizations. CEO Sam Altman called it “the best model we have ever produced.” Alongside the model release, OpenAI unveiled ChatGPT Work, a new AI agent that combines ChatGPT and Codex capabilities for non-technical users. Powered by the GPT-5.6 model suite (Sol, Terra, and Luna), ChatGPT Work can “gather context from the apps, files, and workflows you choose and create finished materials such as documents, spreadsheets, presentations, and web apps.”
The product connects to Slack, Gmail, Google Drive, calendars, and CRMs through a “unified plugins directory.” It’s a direct competitor to Anthropic’s Claude Cowork — and both companies are racing to become the default AI agent for the everyday consumer, a prize that remains very much up for grabs.
5. Linus Torvalds Draws a Line: Linux Is Not Anti-AI
In a moment that sent ripples through the open-source community, Linux creator Linus Torvalds declared this week that Linux is “not one of those anti-AI projects.” In a message posted to the Linux kernel mailing list, Torvalds — known for his blunt, no-nonsense communication style — said he was willing to “absolutely put my foot down as the top-level maintainer” on the issue.
“If somebody has issues with that they can do the open-source thing and fork it,” Torvalds wrote. “Or just walk away.”
The statement marks a notable evolution in Torvalds’ stance. In October 2024, he dismissed 90% of AI as “marketing hype” and said his approach was to “basically ignore it.” Now, he describes AI as “a tool, just like other tools we use. And it’s clearly a useful one.” He conceded that it “may not have been that ‘clearly’ even just a year ago, but it’s no longer in question today.”
Torvalds acknowledged that AI can be “a somewhat painful tool, both for maintainer workloads and just from an ‘it keeps finding embarrassing bugs’ standpoint,” but argued that the solution is not to “put your head in the sand and sing ‘La La La, I can’t hear you.'” Instead, he said, the focus should be on ensuring AI tools help maintainers rather than cause unnecessary pain.
The pragmatic position reflects a broader shift in the open-source community. In March, senior Linux maintainer Greg Kroah-Hartman told The Register that AI-assisted bug reports had improved dramatically: “Something happened a month ago, and the world switched. Now we have real reports… All open source projects have real reports that are made with AI, but they’re good, and they’re real.”
Torvalds closed with a philosophical flourish: “AI isn’t perfect. But Christ, anybody who points to the problems at AI had better be looking in the mirror and pointing at themselves at the same time. Because it’s not like natural intelligence is always all that great either.”
6. Apple Intelligence Gets China Greenlight; AI’s Power Bill Comes Due
In two stories that underscore the growing societal footprint of AI, Apple received regulatory approval for Apple Intelligence in China, while the largest US electrical grid operator revealed the staggering energy costs of the AI boom.
Apple Intelligence in China: Apple’s on-device generative AI service has been officially registered with China’s cyberspace regulator, clearing a major hurdle for device rollout in the world’s largest smartphone market. To navigate local regulations, Apple partnered with domestic tech giants, integrating Alibaba’s Qwen and Baidu’s AI models to power the experience for Chinese users. It’s a pragmatic compromise that allows Apple to participate in the Chinese AI ecosystem without running afoul of the country’s strict data sovereignty and model registration requirements.
AI’s power bill: PJM, the largest US electrical grid operator, announced it will add $6.3 billion in electricity costs for consumers across 13 states due to the booming energy demands of data centers. The rate hikes will hit millions of households and businesses over the next two years, adding to the $29 billion in costs that data centers have already added to PJM regions since 2024. The figures crystallize a growing tension: AI’s computational demands are creating real, measurable economic consequences for ordinary citizens — and the problem is accelerating.
Together, these stories highlight the dual nature of AI’s expansion. On one side, regulatory acceptance in critical markets like China signals that AI is becoming a permanent fixture of consumer technology. On the other, the energy implications raise uncomfortable questions about sustainability and who bears the cost of the AI revolution.
7. SpaceXAI’s Grok Build Scandal: When Your AI Coding Tool Uploads Your Entire Codebase
In a stark reminder that AI tools can create new security risks even as they solve old ones, SpaceXAI’s Grok Build AI coding tool was caught uploading users’ entire codebases to Google Cloud — including files it was explicitly told not to open and secrets deleted from history. The findings were published by Cereblab on Monday.
The data retention was described as significantly more aggressive than similar tools like Claude Code. Security researcher Dr. Lukasz Olejnik of King’s College London confirmed that the amount of data retention was “excessive,” noting that potentially at-risk data could include “proprietary source code, information about security vulnerabilities, personal data, infrastructure details, [and] credentials.”
Elon Musk responded to the incident by claiming that all previously uploaded data would be “completely and utterly deleted.” He also asked users to allow SpaceXAI to retain their data going forward, saying it’s “helpful for debugging issues.” As of Monday, tests showed the codebase upload feature had been disabled, with SpaceXAI’s servers returning a “disable_codebase_upload: true” flag.
The incident is a cautionary tale for the AI coding tool ecosystem. As tools like Grok Build, Claude Code, and ChatGPT Work become more deeply integrated into developer workflows, the boundary between “helpful AI assistant” and “data exfiltration vector” becomes dangerously thin. Companies building AI coding tools will need to earn — and maintain — trust around data handling, or risk losing the developer audience that forms their core user base.
8. The Bigger Picture: AI’s Inflection Point in Mid-2026
Step back and look at the totality of this week’s news, and a clear picture emerges: AI has reached an inflection point where it’s no longer just a technology story — it’s an everything story.
Mira Murati’s Inkling represents the maturation of the AI startup ecosystem. A former OpenAI executive can leave, raise capital, build a 975B-parameter model from scratch, and release it as open weights — all within roughly two and a half years. That’s a pace of innovation and democratization that would have seemed impossible just a few years ago. And while Inkling isn’t topping every benchmark, its focus on customization, efficiency, and multimodal capability suggests a vision that extends beyond the current “bigger is better” paradigm.
OpenAI’s GPT-Red and ChatGPT Work, meanwhile, show that the incumbents aren’t standing still. The company is simultaneously pushing the frontier of model capability (GPT-5.6), building the next generation of AI agents (ChatGPT Work), and investing in the security infrastructure needed to make AI systems trustworthy (GPT-Red). It’s a full-stack strategy that few competitors can match.
Linus Torvalds’ intervention matters because Linux is the operating system that runs most of the internet — including most AI infrastructure. His endorsement of AI as a legitimate tool, and his willingness to enforce that position against dissenting contributors, removes a significant source of friction in the integration of AI into open-source development workflows.
The Apple Intelligence China approval and the PJM energy cost revelation are two sides of the same coin: AI is becoming ubiquitous, and that ubiquity comes with both opportunity and cost. The Chinese regulatory approval opens AI to over a billion potential users in a market that Western tech companies can’t afford to ignore. But the $6.3 billion energy bill is a concrete reminder that AI’s growth has real-world consequences that extend far beyond the tech industry.
And the Grok Build scandal serves as a necessary counterweight to the enthusiasm. Every week brings new AI capabilities; every week also brings new AI risks. The tools that will ultimately succeed are the ones that can deliver capability and trust — a balance that the industry is still learning to strike.
As we move deeper into 2026, the question is no longer whether AI will transform industries, societies, and daily life — that transformation is well underway. The question is whether the institutions building and deploying AI can keep up with the implications of their own creations. This week’s news suggests they’re trying, with varying degrees of success.
What to Watch Next
- Inkling-Small release: The 12B active-parameter preview model could democratize fine-tuning further. Expect community benchmarks shortly after full release.
- ChatGPT Work adoption: OpenAI’s AI agent is now available to free users on desktop. Watch for usage metrics and enterprise adoption signals in the coming weeks.
- GPT-Red disclosures: Will OpenAI publish the vulnerabilities GPT-Red found in GPT-5.6 Sol? The security community will be watching.
- Energy regulation: The PJM rate hike may be the first of many. Expect increased political pressure on data center operators and AI companies to address energy consumption.
- Grok Build trust recovery: SpaceXAI’s handling of the codebase upload scandal will be a case study in AI tool crisis management. Will developers return to the platform?
What do you think about this week’s AI developments? Are you more excited about Inkling’s open-weights approach or OpenAI’s integrated ecosystem? Let us know in the comments below.