Former OpenAI CTO Mira Murati Launches Thinking Machines' Inkling: Open-Weights Multimodal AI Model with 975 Billion Parameters and Self-Training Demo
Releasing yet another powerful language model is no longer enough to stand out. Thinking Machines, the startup founded by former OpenAI technical director Mira Murati, has chosen a different approach to attract developers. The company unveiled Inkling, its first open-weights model that can be freely used and adapted for specific tasks.
Inkling is a multimodal model capable of processing text, images, and audio without separate processing modules for each modality. It is built on a mixture-of-experts architecture containing 975 billion parameters, of which only 41 billion are active at any given time. This design significantly reduces computational costs. The model supports a maximum context length of 1 million tokens and was trained on a dataset of 45 trillion tokens that included text, images, audio, and video. Alongside the main model, the company released a preliminary Inkling Small version with 12 billion active parameters optimized for lower-cost and faster deployment.
One of Inkling’s standout features is the ability to regulate reasoning depth. Developers can choose how many computational resources the model should allocate to solving a problem. For simple queries the system responds faster and consumes fewer tokens, while for complex tasks it can increase internal computation. According to the company, in certain programming scenarios Inkling uses approximately three times fewer tokens than several other open models while maintaining comparable quality.
Thinking Machines acknowledges that Inkling does not yet aim to be the strongest model on the market. Closed models from OpenAI, Anthropic, and Google continue to lead most comprehensive benchmarks, while Chinese models remain ahead in certain disciplines. Instead of competing for top rankings, the company focused on creating a versatile foundation for subsequent training and customization for enterprise use cases.
To support further development, Thinking Machines offers its own Tinker platform. Through this platform, developers can fine-tune Inkling on their own data without building complex infrastructure themselves. To demonstrate the platform’s capabilities, the company conducted an unusual experiment: Inkling was tasked with training itself. The model autonomously created a training task, executed the process via Tinker, evaluated the results, and switched to the updated version. In the demonstration, it was given the unusual objective of learning to answer questions while completely avoiding the use of one letter of the English alphabet.
The company has also made the model weights available on Hugging Face and added support for popular inference frameworks including Transformers, vLLM, SGLang, and llama.cpp. Inkling is designed not only for cloud services but also for continued training, creation of specialized assistants, and development of autonomous agents capable of executing complex sequences of actions.
The release of Inkling represents an important development for the Western open-weights AI community. After Meta reduced its activity in this area and many organizations began turning to Chinese models, developers now have another major Western-origin solution that can be freely run, modified, and adapted to their own tasks thanks to its open weights.
Related articles
Yandex Black October: How Architectural Flaws and Capacity Mismanagement Turned Data Center Losses Into a 11-Day Fintech Coma
In October 2026 three Yandex data centers in Central Russia were physically destroyed within days, triggering the collapse of Yandex Cloud and multiple core services. Between the first and third site failures engineers had 74 hours to restore quorum and rebalance workloads yet failed to do so because Control Plane components remained tied to the lost zone and spare capacity had been eliminated to cut costs. Core banking and BNPL systems entered Read-Only mode after losing quorum, forcing a complete stop-the-world reconciliation that lasted until 22 October. Backups stored inside the same regional S3 buckets proved unusable, resulting in a seven-day RPO and primary-key collisions after restoration. The incident exposed that Metro-Cluster designs had been incorrectly treated as full Disaster Recovery and that aggressive AP microservice patterns created logical split-brain conditions outside the database layer.
Astra Group Releases Low-Code Astra AI Platform for Building Corporate AI Agents in Closed Environments
Astra Group has launched the Astra AI Platform, a low-code constructor designed for creating corporate AI applications that operate entirely within closed circuits without reliance on foreign clouds. The platform targets small and medium businesses as well as large corporations, enabling rapid deployment of AI secretaries, technical support assistants, and other agents in days rather than months. It forms part of the broader Astra AI ecosystem and includes two main constructors: one for machine learning model training and deployment, and another for LLM-based language assistants, document search, and process automation. Users can assemble scenarios from visual blocks, debug them in real time, or extend functionality with custom Python code when needed. The solution supports RAG-based knowledge bases, integration with external tools such as CRM and ERP systems, and the coordination of multiple agents for distributed task handling. One demonstrated use case involves processing ten pages of engineering drawings and specifications in five minutes to prepare data for downstream systems. The product is already listed in the Russian domestic software registry and available to customers.
Multi-Tenancy Isolation in VictoriaMetrics, VictoriaLogs and Grafana: Building Client Separation in a Shared Stack
Voltir runs a single shared monitoring stack consisting of VictoriaMetrics, VictoriaLogs, Grafana and vmalert for all customers instead of deploying separate instances per tenant. Isolation relies on numeric account IDs enforced at every layer: nginx validates tokens against the platform API before proxying writes, Grafana data sources hard-code tenant accounts, and vmalert rules receive automatic account filters. The team discovered that a single malformed rule or template could break alerting for every tenant because vmalert loads the entire rules directory at once. They also found that Grafana automatically created personal organizations granting Org Admin rights, accumulating hundreds of orphaned organizations and users that consumed disk space and resources. A dedicated cleanup module plus hourly reconciliation now removes rules, organizations, series and users when a tenant is deleted, while isolation tests verify both data and API objects across tenants.
The Real Costs of Using Claude and GPT for Russian Businesses Exposed
Accessing generative AI models like Claude and GPT has become straightforward for Russian companies, yet calculating their true expenses remains complex. Hidden factors such as intermediary commissions, currency conversion rates, taxes, and the often-overlooked costs of model reasoning tokens can double or triple initial estimates. Founder Vitaly Nagay of the Heli service analyzed these financial and organizational risks for Russian firms relying on foreign AI providers. In one example, a support department processing 20,000 monthly queries sees costs rise from $240 to $480 when reasoning tokens are included. Nagay recommends evaluating providers by dividing actual ruble charges by the model's official dollar pricing to reveal hidden markups. Additional concerns involve data control, document compliance, and access history when employees use personal subscriptions. Companies are advised to verify billing details, data storage policies, and issue dedicated API keys with spending limits before procurement.