Thinking Machines Lab has released Inkling, a 975 billion-parameter open-weight multimodal model that emphasizes customization over frontier benchmark leadership, but its established rivals and steep deployment requirements leave the business case unproven.
Mira Murati’s heavily financed startup now has a foundation model; the harder test is whether customers value control enough to accept its performance and infrastructure trade-offs.
Thinking Machines Lab released Inkling on July 15 as a general-purpose model with downloadable weights that developers can modify. The company’s model card describes a 975 billion-parameter, 41 billion-active-parameter system that accepts text, images and audio and returns text. The Apache 2.0 release supports a context window of up to 1 million tokens when run from the open weights; access through the company’s Tinker platform is limited to 256,000 tokens.
The release turns Thinking Machines’ enterprise pitch into a product choice. Customers can obtain the weights through Hugging Face, fine-tune Inkling through Tinker or use a third-party inference provider. A report on its enterprise strategy describes the bet as customization producing better performance at lower cost. That remains a proposition, not a result demonstrated across customer deployments.
Open weights do not amount to full disclosure of the model’s development. The model card says training data came from public sources, third parties or synthetic generation and augmentation, but it does not identify individual datasets. Developers can inspect and modify the released parameters without receiving a complete inventory of the material that produced them.
Nor is customization unique to Inkling. OpenAI offers customizable open-weight models, while Amazon and Microsoft sell services for fine-tuning models on customer data, according to a report on the launch. Open weights can give a deployer more choice over hosting and modifications, but Inkling enters a market with both downloadable substitutes and managed customization platforms.
Thinking Machines says it trained Inkling from scratch rather than modifying another company’s checkpoint. The company also said, in details carried by the same launch report, that Inkling’s architecture drew on DeepSeek-V3 and that its post-training used data generated by Moonshot AI’s Kimi K2.5. Those statements can coexist: ownership of the starting weights does not mean the design and post-training inputs were developed without outside technical influence.
That qualification matters because the release arrives amid complaints from leading U.S. labs that Chinese developers use American model outputs for distillation. Here, the direction of influence is partly reversed: a richly funded San Francisco company used architectural ideas and synthetic outputs from Chinese systems. The evidence does not establish dependence on any single rival, but it does show that open model development cuts across the national competition often used to frame it.
Inkling scored 41 on the Artificial Analysis Intelligence Index, three points above Nvidia’s Nemotron 3 Ultra. That makes it the highest-scoring open-weight release from a U.S. lab in the third-party benchmark analysis. It also averaged 25,000 output tokens per index task, compared with 43,000 for GLM-5.2, 38,000 for Kimi K2.6 and 37,000 for DeepSeek V4 Pro under the evaluator’s named configurations.
The same evidence limits any broader claim. Inkling ranks below leading closed products from Anthropic and OpenAI and below several Chinese models overall. Its results also shift by task: it beat Kimi K2.6 and DeepSeek V4 Flash on two cited agentic evaluations, while the company’s own table shows losses to different rivals on multiple reasoning, coding, factuality, vision and audio tests. On Artificial Analysis’s omniscience evaluation, Inkling posted 40% accuracy and a 63% hallucination rate.
Thinking Machines also used Inkling to fine-tune itself. The company said that process made the model’s visible chain of thought more concise without changing the final response, according to an account of the release. The experiment demonstrates an internal use of the fine-tuning system, but it does not show that customers will reproduce useful gains on proprietary tasks or at lower total cost.
Running the full BF16 checkpoint requires at least 2 TB of aggregated GPU memory, configured as eight Nvidia B300 GPUs or 16 H200 GPUs. The quantized NVFP4 checkpoint reduces the minimum to 600 GB, using four B300s or eight H200s in the configurations specified by Thinking Machines. Deployment also requires an inference framework and its dependencies. The weights are open; operating the largest version remains a cluster-scale undertaking.
Hosted access therefore stays central to the economics. Tinker pricing at a 64,000-token context is $1.87 per million input tokens, $0.374 per million cached tokens and $4.68 per million output tokens. At 256,000 tokens, each rate doubles. Organizations can avoid exclusive reliance on Thinking Machines by self-hosting or choosing a third-party provider, but they do not avoid the cost and concentration of advanced GPU infrastructure.
The company’s safety assessment similarly transfers work to deployers. Thinking Machines says Inkling is materially below public frontier models in loss-of-control capabilities and does not add a material capability uplift beyond existing open-weight systems. It nevertheless reports occasional compliance with harmful role-play and indirectly framed prompts. Its recommendations include input and output filtering, rate limits, monitoring, domain-specific validation and human oversight—controls customers must test after any fine-tuning.
Murati founded Thinking Machines in February 2025 after serving as OpenAI’s chief technology officer and briefly as its chief executive. Early leaders included OpenAI co-founder John Schulman and former OpenAI vice president Lilian Weng. Before Inkling, the startup released Tinker and previewed an interaction model. The launch report also said several senior employees left for Meta and OpenAI in early 2026 and questioned Murati’s leadership, adding an execution concern to the technical test.
The financial expectations arrived before the foundation model. Thinking Machines raised a $2 billion seed round in 2025 at a $12 billion post-money valuation, with Nvidia, AMD, Andreessen Horowitz and Jane Street among its backers. Nvidia supplied the latest AI infrastructure used to develop Inkling and is also an investor. The enterprise-strategy report says the company also signed a multibillion-dollar Google Cloud agreement.
That funding makes Inkling more than a technical debut: it is the first production foundation-model evidence against a valuation established before the company had released a model or product. Yet it cannot by itself validate that price. The release follows existing enterprise tools, trails stronger general-purpose systems and relies on infrastructure supplied by companies with their own leverage over AI deployment.
Thinking Machines is previewing Inkling-Small, whose weights it says will follow after testing, and says more powerful successors are in development. The immediate question is measurable: can customers produce enough improvement on specialized work to offset fine-tuning, inference, safety and hardware costs when rival open models and managed services are already available? Evidence from repeat deployments, total operating costs and task-specific gains—not the U.S.-only leaderboard title—will determine whether Inkling creates durable enterprise leverage or merely another route to customization.
Shanghai AI Laboratory has released Intern-S2-Preview-397B as downloadable weights and a hosted API, but its own 35B alternative offers the same context window and a far smaller disclosed scale while the larger model still lacks auditable comparative scores, pricing and physical-validation results.
Mixfont's free Decoy Font gives each glyph a sharp decoy and a blurred intended letter, creating a hurdle for pixel-based readers while leaving selectable text, known-image techniques and accessibility costs outside its protection.
Zhipu reportedly reached $1 billion in annual recurring revenue in July, roughly four times a March estimate, but the unconfirmed run rate is not annual sales and still sits far ahead of recognized cloud revenue while margins remain thin.
Anthropic’s 20-year lease gives TeraWulf a customer for 401 megawatts of future AI infrastructure, but rent starts only after delivery and the proposed utility deal passes power and grid costs to TeraWulf, leaving financing and project margin unresolved.
OpenAI has put conversations and Projects back in its redesigned ChatGPT desktop app and enabled cloud Work threads to move across devices, correcting the launch's biggest usability failures without merging local Work or Codex histories.
Twenty-nine countries signed an agreement creating WAICO as an independent intergovernmental organization, while China paired the launch with capacity-building offers that are not yet confirmed as WAICO programs.
Meta reportedly plans to put departing AWS compute executive Dave Brown to work on its data-center buildout, adding hyperscale operating experience while leaving any customer-facing cloud business conditional and undefined.
Moonshot AI has made Kimi K3 available through its apps and API, pairing a 2.8-trillion-parameter architecture with early frontier-level results, but the model's open-weight claim cannot be tested until its weights and technical report arrive.
CIA Director John Ratcliffe said US intelligence is consistent with an estimate that Russian recruits last 20 to 30 minutes on Ukraine’s battlefield, but the public trail leads to an unsourced claim about assault troops and does not establish a representative average.
China has paired a five-year AI training offer for developing countries with cooperation centers, a weather-warning rollout and a new 29-country organization. The package gives Beijing a platform for influence, but no budget, selection rules or delivery timetable has been published.
Federal intervention changed the launches of Anthropic's Claude Fable 5, Claude Mythos 5 and OpenAI's GPT-5.6, but through different legal and voluntary channels. The result is real government leverage over early customers without a common security threshold or durable review process.
A proposal developed with Treasury Secretary Scott Bessent would put an independent AI regulator under SEC oversight, but President Trump had not reviewed it and its tests, funding and enforcement authority remained unresolved.
San Francisco has demanded that Apple and Google cut off 13 apps capable of producing nonconsensual sexual deepfakes. The threatened case could test California’s new liability rules, but the confidential app list and the law’s “primary purpose” definition leave a central question unresolved.
Anthropic will keep Claude Fable 5 inside Max and Team Premium plans from July 20 at 50% of their usage limits, while Pro and Team Standard customers move to separately billed credits. The durable change is who gets bundled access, not a larger allowance for top-tier users.
HKT plans a 3.2 Tbps data-centre route from Lok Ma Chau to Tseung Kwan O by the end of 2026, but has not disclosed customer bandwidth, price, project cost or end-to-end latency, while Sandy Ridge's developer expects to start within 42 months of its land award.
Apple briefly topped Nvidia during Friday trading before Nvidia closed about $6 billion ahead. The filings show why that ranking is a weak AI verdict: Apple’s operating momentum predates its new Siri, while Nvidia’s rapid growth comes with concentrated customers and large forward commitments.
Anthropic has proposed paying Meta as much as $10 billion over two years for AI computing capacity, but early exit rights, undisclosed capacity and Meta's own reliance on outside infrastructure leave the economics—and the case for a durable Meta cloud business—unresolved.
Barracuda says it detected more than one million retail-themed phishing emails using hidden filler text since April, but its published research does not show how many passed security controls or whether AI-generated filler performs better than older variants.
A federal judge denied emergency relief to 26 Meta workers challenging allegedly AI-assisted layoff selections, finding no near-term irreparable harm while leaving the discrimination claims, a preliminary-injunction request and the role of Meta's internal tools unresolved.
Google says it fixed a Gemini flaw that let someone with physical access to a locked Android phone send messages and reconnect apps without the expected PIN check, but it has not identified the affected devices, delivery channel, or software version that would let users verify the repair.