Articles in the AI-Models section on Beckmann.https://beckmann.aien-UShttps://beckmann.ai/en/ai-models/2026-07/thinking-machines-inkling-open-ai-modelhttps://beckmann.ai/en/ai-models/2026-07/thinking-machines-inkling-open-ai-modelThinking Machines Lab released Inkling, its first open language model with 975 billion parameters, of which only 41 billion are active, on July 15, 2026. The model processes text, image, and audio natively and can be tailored to specific fields via the Tinker platform. Founder Mira Murati positions Inkling not as the most powerful model, but as a customizable base for companies.Thu, 16 Jul 2026 07:15:55 GMTAI-ModelsBrian Beckmannhttps://beckmann.ai/en/ai-models/2026-07/prismml-bonsai-27b-iphonehttps://beckmann.ai/en/ai-models/2026-07/prismml-bonsai-27b-iphoneCalifornia startup PrismML has compressed its Bonsai 27B model, with 27 billion parameters, down to 3.9 gigabytes and made it run on iPhones. Quantization to as little as 1.58 bits per weight makes this possible. The company says up to 95 percent of the original performance is retained. Apple is evaluating the technology for its own devices, according to founder Babak Hassibi.Thu, 16 Jul 2026 06:22:21 GMTAI-ModelsBrian Beckmannhttps://beckmann.ai/en/ai-models/2026-07/robbyant-lingbot-va-20-robot-aihttps://beckmann.ai/en/ai-models/2026-07/robbyant-lingbot-va-20-robot-aiOn July 10, 2026, the Ant Group robotics unit Robbyant released its latest foundational model LingBot-VA 2.0 as open source. The system controls robotic arms in simulation tests with a success rate of 93.6 percent and significantly accelerates the response time compared to the previous version. The code is freely accessible on GitHub.Sat, 11 Jul 2026 12:44:08 GMTAI-ModelsBrian Beckmannhttps://beckmann.ai/en/ai-models/2026-07/mistral-robostral-navigate-roboticshttps://beckmann.ai/en/ai-models/2026-07/mistral-robostral-navigate-roboticsMistral AI unveiled its first model for robot navigation, Robostral Navigate, on July 8, 2026. The 8-billion-parameter model navigates robots solely using a single RGB camera and natural language instructions through complex indoor and outdoor spaces, entirely without Lidar or depth sensors. According to Mistral, it achieves a 76.6 percent success rate on the R2R-CE benchmark in unknown environments, surpassing comparable single-camera systems by 9.7 percentage points. The model is part of Mistral's strategy to position itself as a European provider of industrial 'Physical AI' following the acquisition of the Vienna-based physics AI startup Emmi AI in May 2026. Mistral has not yet released public model weights, a technical paper, or a pricing model.Fri, 10 Jul 2026 10:18:04 GMTAI-ModelsBrian Beckmannhttps://beckmann.ai/en/ai-models/2026-07/openai-chatgpt-work-gpt56-agent-enhttps://beckmann.ai/en/ai-models/2026-07/openai-chatgpt-work-gpt56-agent-enOn July 9, 2026, OpenAI introduced ChatGPT Work: an AI agent that independently plans and executes multi-step tasks over hours, collects context across apps (including Slack, Salesforce, Google Drive, Jira), and creates finished documents, spreadsheets, presentations, and web applications from it. At the same time, OpenAI launched the new model family GPT-5.6 in three price tiers: Sol (flagship, $5 / $30 per million tokens), Terra ($2.50 / $15), and Luna ($1 / $6). ChatGPT Work is now available for Pro, Enterprise, and Edu users, with Plus and Business following within a few days; the desktop app for Mac and Windows is open to all plans including Free starting today.Thu, 09 Jul 2026 19:20:03 GMTAI-ModelsBrian Beckmannhttps://beckmann.ai/en/ai-models/2026-07/meta-muse-image-instagram-photos-privacyhttps://beckmann.ai/en/ai-models/2026-07/meta-muse-image-instagram-photos-privacyOn July 7, 2026, Meta launched Muse Image, a new AI image model from Meta Superintelligence Labs, available in Meta AI, Instagram Stories (USA), and WhatsApp. The model allows public Instagram photos to be integrated into AI-generated images via @-mention - activated by default, without opt-in. Following criticism from privacy organizations like Public Citizen, media outlets like Malwarebytes and The National show how users can disable the feature through Instagram settings. Business-wise, Muse Image primarily targets advertisers: it complements Meta's automated Advantage+ system, which according to Meta generates around $60 billion in annual revenue.Thu, 09 Jul 2026 18:12:03 GMTAI-ModelsBrian Beckmannhttps://beckmann.ai/en/ai-models/2026-07/grok-45-spacexai-coding-modelhttps://beckmann.ai/en/ai-models/2026-07/grok-45-spacexai-coding-modelSpaceXAI introduced Grok 4.5 on July 8, 2026, a model focused on coding and agentic work. The price is $2 per million input and $6 per million output tokens, significantly undercutting Anthropic's Opus 4.8. Elon Musk describes it as "Opus-class, but faster and cheaper." In the four benchmarks published by SpaceXAI itself, Grok 4.5 beats Opus 4.8 in two and loses in two; Claude Fable 5 leads all four. The numbers have not yet been independently verified. The model was trained in conjunction with the coding tool Cursor and is initially not available in the EU.Thu, 09 Jul 2026 17:16:40 GMTAI-ModelsBrian Beckmannhttps://beckmann.ai/en/ai-models/2026-07/meta-muse-spark-11-price-warhttps://beckmann.ai/en/ai-models/2026-07/meta-muse-spark-11-price-warOn July 9, 2026, Meta released a new multimodal reasoning model, Muse Spark 1.1, and for the first time launched a paid developer API for its own frontier model: $1.25 per million input tokens, $4.25 per million output tokens. According to Meta's own benchmarks, the model outperforms Claude Opus 4.8 and GPT-5.5 in agent and tool use tasks (including MCP Atlas, JobBench, Humanity's Last Exam), but falls behind in pure coding (SWE-Bench Pro) and multimodal tasks. The public preview is initially limited to developers in the USA, with an EU launch still pending. This step confirms Meta's shift away from the open Llama strategy in favor of closed models under Meta Superintelligence Labs, initiated in April 2026.Thu, 09 Jul 2026 00:00:00 GMTAI-ModelsBrian Beckmannhttps://beckmann.ai/en/ai-models/2026-07/tencent-hy3-moe-modelhttps://beckmann.ai/en/ai-models/2026-07/tencent-hy3-moe-modelTencent has officially released Hy3 as an open-source language model with 295 billion total parameters and 21 billion active parameters. The model, built on a mixture-of-experts architecture, aims to achieve performance parity with models that have 2-5x more parameters and has been released under the commercially friendly Apache-2.0 license. The rapid development cycle – from infrastructure overhaul in January to official release in July 2026 – indicates a new production understanding in the AI industry.Mon, 06 Jul 2026 22:36:33 GMTAI-ModelsBrian Beckmann