Articles in the Open Source section on Beckmann.https://beckmann.aien-UShttps://beckmann.ai/en/security/2026-08/mistral-shieldstral-ai-guardianhttps://beckmann.ai/en/security/2026-08/mistral-shieldstral-ai-guardianMistral launched Shieldstral, an open AI safety model with three billion parameters, on August 4, 2026. It evaluates text and images against freely formulated rules, without companies needing to retrain it. On a single 16-gigabyte graphics card, it reportedly matches the accuracy of guardian models up to seven times larger.Fri, 07 Aug 2026 06:29:34 GMTSecurityEuropeOpen SourceBrian Beckmannhttps://beckmann.ai/en/ai-policy/2026-08/white-house-ai-framework-excludes-open-sourcehttps://beckmann.ai/en/ai-policy/2026-08/white-house-ai-framework-excludes-open-sourceFive US tech companies discussed the government's completed but unpublished AI review regime at the White House on August 4, 2026. Only closed top models from OpenAI, Google, Anthropic, Meta, and Microsoft will have to undergo a 30-day review before release. Providers that disclose their model weights are fully exempt from the rule.Thu, 06 Aug 2026 14:36:29 GMTAI-PolicyOpen SourceBrian Beckmannhttps://beckmann.ai/en/security/2026-07/aisi-cyber-gap-open-ai-modelshttps://beckmann.ai/en/security/2026-07/aisi-cyber-gap-open-ai-modelsAn analysis by the AI Security Institute from July 17, 2026 shows: Chinese open-weight systems like GLM-5.2 and DeepSeek V4-Pro are closing the gap in attack capabilities against proprietary top models. The gap shrank within a year from six-to-ten months to four-to-seven months. An independent benchmark by Semgrep confirmed the trend already in June.Sat, 18 Jul 2026 11:19:13 GMTSecurityOpen SourceBrian Beckmannhttps://beckmann.ai/en/ai-in-practice/2026-07/linux-kernel-ai-torvalds-sashikohttps://beckmann.ai/en/ai-in-practice/2026-07/linux-kernel-ai-torvalds-sashikoLinux creator Linus Torvalds has confirmed on the kernel mailing list that the kernel will actively rely on AI tools such as the code review tool Sashiko. Critics point to false positives, a halted automatic Reviewed-by tag and the recommendations of the Software Freedom Conservancy – Torvalds counters that decisions are made on technical, not ideological criteria.Fri, 17 Jul 2026 11:07:21 GMTAI in PracticeLinuxOpen SourceSashikoBrian Beckmannhttps://beckmann.ai/en/ai-models/2026-07/kimi-k3-moonshot-open-source-catch-uphttps://beckmann.ai/en/ai-models/2026-07/kimi-k3-moonshot-open-source-catch-upMoonshot AI has today released Kimi K3: a 2.8 trillion parameter model with one million token context, which is said to perform at Opus-4.8 level in benchmarks. The open model shows that local and open-source AI development is catching up to the global top tier.Thu, 16 Jul 2026 07:20:00 GMTAI-ModelsOpen SourceBrian Beckmannhttps://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-ModelsOpen SourceBrian Beckmannhttps://beckmann.ai/en/ai-economy/2026-07/nous-research-hermes-valuationhttps://beckmann.ai/en/ai-economy/2026-07/nous-research-hermes-valuationNous Research, provider of the open AI agent Hermes, is reportedly negotiating a funding round of at least $75 million. The targeted valuation of $1.5 billion would be about half above the Series A from a year ago. Robot Ventures and Union Square Ventures are considered the backers of the round. No confirmation from the involved companies has been received so far.Wed, 15 Jul 2026 16:28:44 GMTAI-EconomyStock PricesOpen SourceBrian Beckmannhttps://beckmann.ai/en/research/2026-07/soofi-s-german-open-ai-modelhttps://beckmann.ai/en/research/2026-07/soofi-s-german-open-ai-modelA German research consortium has released Soofi S, an open language model with 31.6 billion parameters for German and English. It was trained on Deutsche Telekom's Industrial AI Cloud in Munich, funded with around 20 million euros from the federal economics ministry. The developers report top scores among open models on combined German-English benchmarks, though these figures have not been independently verified.Wed, 15 Jul 2026 13:03:12 GMTResearchGermany & AIOpen SourceBrian 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-ModelsRobotsOpen SourceBrian Beckmannhttps://beckmann.ai/en/ai-economy/2026-07/ollama-series-b-65-millionhttps://beckmann.ai/en/ai-economy/2026-07/ollama-series-b-65-millionThe open-source provider Ollama announced a Series B funding round of $65 million on July 9, 2026, led by investor Theory Ventures. This brings the total capital raised by the company, founded in 2023, to $88 million. Ollama allows developers to run open AI models locally on their own machines or optionally via a cloud service. According to company information, around 8.9 million developers now use the software monthly, compared to about 4.45 million in early 2026. The deal is part of a series of financings that show investors are increasingly focusing on infrastructure around open rather than closed AI models.Fri, 10 Jul 2026 03:24:52 GMTAI-EconomyOpen SourceBrian 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-ModelsOpen SourceBrian Beckmannhttps://beckmann.ai/en/ai/2026-07/mistral-mensch-warning-closed-ai-modelshttps://beckmann.ai/en/ai/2026-07/mistral-mensch-warning-closed-ai-modelsMistral CEO Arthur Mensch warns companies in a LinkedIn post against relying on closed AI models, as providers increasingly store customer data and gain insight into business processes – some have allegedly used this, according to Mensch, to target successful customers as competitors, although he provides no evidence for this. Palantir CEO Alex Karp expressed similar criticism shortly before and published an 'AI-Sovereignty' manifesto. Both statements also clearly serve their own business interests. An experiment by Bridgewater and Thinking Machines Lab, in which an open model was retrained with expert data, partially supports the perspective: it achieved higher accuracy and lower costs in financial tasks than large frontier models, although not independently verified.Mon, 06 Jul 2026 07:54:02 GMTAIEuropeOpen SourceBrian Beckmannhttps://beckmann.ai/en/ai/2026-07/pxpipe-image-tokens-claude-code-costshttps://beckmann.ai/en/ai/2026-07/pxpipe-image-tokens-claude-code-costsThe open-source tool pxpipe renders extensive text inputs for Claude Code – system prompts, tool documentation, older chat history – as PNG images because Anthropic charges flat rates based on pixel size rather than text content. According to developer Steven Chong, this reduces the overall bill by an average of 59 to 70 percent. The method is lossy: exact strings like hashes can be misrepresented from images without any detectable error. By default, pxpipe only supports Claude Fable 5 and GPT 5.6, as Opus and GPT-5.5 models read image content measurably worse. The basic idea is not new and builds on DeepSeek's OCR model for optical context compression.Sun, 05 Jul 2026 01:49:45 GMTAIOpen SourceBrian Beckmann