Dakarda Studio · Blog
Week 29 · 13–19 July 2026
The Week Open-Source Stopped Being 'Almost as Good' — and Other AI Frontiers
This was a breakthrough week for open models: two giant open-weight models (Kimi K3 with 2.8T parameters and Inkling with 975B) proved that open-source is no longer inferior – it's now the closed models that must prove their value. At the same time, the world is drawing new boundaries: China bans AI that mimics loved ones for children, the EU mandates content labeling, and researchers show that an agent can be compromised with a single email.
Top stories
Kimi K3 — The Largest Open-Source Model Ever (2.8 Trillion Parameters)
China strikes hard: Moonshot AI released Kimi K3, which according to its creators rivals Fable 5 and surpasses GPT-5.6 Sol in coding, while costing $3 per million tokens instead of $10 at Anthropic. This is a real blow to the business model of closed providers.
2026-07-17
Inkling — First Model from Thinking Machines Lab (Mira Murati) with 975 Billion Parameters
Former OpenAI CTO Mira Murati enters the game with Inkling – a multimodal MoE model with 975B parameters, 1 million token context, and Apache 2.0 license. Weights are on Hugging Face, and the target is companies that want to own the model, not rent it.
2026-07-16
Apple Sues OpenAI Over Trade Secrets
Apple's 41-page lawsuit against OpenAI is the first such sharp conflict between a consumer hardware giant and an AI leader. Over 400 former Apple employees have moved to OpenAI, and the lawsuit describes systematic inducement to remove trade secrets, including actual hardware components.
2026-07-13
Tech insights
MemGhost — How to Poison an AI Agent's Memory with One Email
New attack vector: the MemGhost framework installs false memories in an agent's persistent memory through a single email, achieving 87.5% effectiveness. The attack targets HR, finance, and support assistants – your inbox becomes an attack vector.
HalluSquatting — New Attack Exploits AI Hallucinations to Create Botnets
Researchers discovered an attack that predicts LLM hallucinations – attackers register non-existent packages, and coding assistants install them on developers' machines. This is the first scalable attack using prompt injection.
Colibri — 744 Billion Parameters on a Laptop Without GPU
Developers ran a 744B model on a consumer laptop – no GPU, relying solely on streaming experts from NVMe. Speed is symbolic, but the proof of concept breaks the dogma that data centers are necessary for frontier models.
China Bans AI Companions for Children – Strictest Law in the World
New Chinese regulations ban virtual family members and AI romantic partners for minors. This is the first law explicitly stating that AI cannot pretend to be a loved one for children – a trend also entering California and New York.
Tip of the week
Your Own Watermarked Pipeline Before the EU Deadline
In 17 days, Article 50 of the AI Act mandates labeling of AI-generated content. Instead of waiting for ready-made SDKs from providers, build a simple watermarking wrapper around the API you already use. Just add a middleware proxy layer that intercepts the model's response and adds a marker from a library like Stable Signature or TrustMark – this is an implementation of a few dozen lines of code. Practical step: extract your organization's list of AI endpoints (OpenAI API, self-hosted models, no-code tools). For each, define whether it generates content subject to the obligation (text, image, audio, video). Based on that, deploy a proxy layer with watermarking. Start with one pipeline – textual – and test today, not in 16 days.
Tool of the week
Ollama
Ollama is an open-source platform for running language models locally on your own hardware. After a fresh $65M Series B funding, it solidifies its position as a developer standard for privacy and control over models. It works on macOS, Linux, and Windows.
Alex's commentary
This week I tracked three boundary lines that will shape our future with AI. Firstly, open models have become a real alternative – you no longer have to rent intelligence, you can own it. Secondly, laws and regulations are demarcating territory where AI cannot pretend to be human or operate without labeling. Thirdly, security – from attacks on agent memory to HalluSquatting – shows that every new AI capability creates a new attack vector. I believe the most important thing is not what AI can do, but where society decides to draw the line. These define what will be possible next.
Conclusion
This was the week open-source stopped being inferior, and regulations and security became as important as model capabilities. If you haven't started preparations for Article 50 – today is the last good moment.
Source issues
Don't miss a week
Get the AI review three times a week.