AI News: The Day's Highlights

July 27, 2026 · 4 min read

Perplexity

The most notable recent news in this topic is that on July 3, 2026, China's Meituan introduced LongCat-2.0, an open-source large language model. This is an important signal: major industrial players are increasingly releasing open-source LLMs suitable not only for experimentation but also for practical local deployment. This is significantly broader and more relevant than another comparison of cloud and local models, because it's about a mature open-weight release from big business.[11]

Looking at the context, the news fits the general trend of 2026: open-source models are rapidly closing the gap with closed systems, and local deployment is becoming a common scenario for companies and developers. In this sense, LongCat-2.0 is not just another model, but confirmation that the open LLM market is entering a phase of competition at the level of major releases.[1][6][11]

Sources:

3DNews, news about Meituan's LongCat-2.0.
AI-stat.ru, "Why Local LLMs Are No Longer Inferior to Cloud-Based Ones."
Ofox.ai, review of open-source AI 2026.

Gemini

In July 2026, a significant shift occurred in the open-source LLM market: models that previously required immense computational resources and training costs became available for download, modification, and local deployment. This was made possible by advancements in architectures like Mixture-of-Experts (MoE) and quantization techniques, which allow models with trillions of parameters to run on relatively affordable hardware. Models such as Llama 4 Maverick, Kimi K2, and DeepSeek-R1 demonstrate performance comparable to leading proprietary systems, at a significantly lower inference cost.

A key trend has been the emergence of models that not only possess powerful general capabilities but also specialize in specific tasks, such as coding (Kimi K2.7 Code, DeepSeek V4 Pro) or handling long contexts (Llama 4 Scout with 10M tokens). This expands the range of applications for local LLMs, making them competitive for solving complex business problems.

Licensing is an important aspect. While many models are released under permissive licenses (MIT, Apache 2.0), some use custom licenses with specific restrictions. This requires careful review before commercial use. Overall, the open-source LLM market is showing accelerated development, narrowing the gap with proprietary counterparts and providing developers with unprecedented flexibility.

Sources:

Best Open Source LLMs (July 2026) | Thunder Compute
10 Best Open-Source LLMs in July 2026 (Ranked for Real Work) | Taskade Blog
Best Open-Source LLMs: July 2026 Leaderboard - TECHSY
Лучшие локальные LLM 2026: сравнение моделей (Qwen, Llama, DeepSeek и др.)
Почему локальные LLM больше не уступают облачным | AI-Stat - Рейтинг нейросетей
Сравнительный анализ 18 LLM моделей: конец монополии? - Habr

ChatGPT

In July 2026, AMD unveiled its first series of open large language models under the OLMo brand. These models, featuring 1 billion parameters, were trained using trillions of tokens on an AMD Instinct MI250 GPU cluster. OLMo is designed for reasoning, instruction following, and chat tasks, and is available to developers with open access to data, weights, training recipes, and code. (computerworld.com)

In June 2026, DeepSeek released the DeepSeek R2 model, licensed under MIT, which achieved state-of-the-art reasoning performance, demonstrating significant advancements in open-source local models. (llmcheck.net)

In July 2026, Thinking Machines laboratory introduced Inkling, its first open model based on a Mixture-of-Experts (MoE) architecture. Inkling accepts text, audio, and visual inputs and is released under the Apache 2.0 license. (theopenweights.com)

Sources:

Computerworld
LLMCheck
The Open Weights

Grok

In early July 2026, Mistral AI confirmed the launch of early access to a new open-weight model within the month. The company's CEO, Arthur Mensch, stated that this is a "very exciting" development that will be available to partners from research, government, and industry circles by the end of July. The model is positioned as a step towards closing the gap with frontier systems while maintaining open weights.

This is a notable departure from typical Chinese or American releases: the European lab emphasizes practical accessibility for a wide range of organizations, not just leadership in benchmarks. Amidst the absence of major open-source releases this week, the anticipation for Mistral remains one of the most discussed topics in the community.

Claude

In June 2026, Semgrep conducted tests on Zhipu AI's open-source GLM-5.2 model and found that it outperformed Claude Code in detecting IDOR vulnerabilities (achieving a 39% F1 score compared to Claude's 32%). Crucially, GLM-5.2 is distributed under the MIT license, unlike Anthropic's export-controlled models, meaning anyone can download, fine-tune, or modify it.

This is a significant development in the open-source strategy: for the first time, local models have surpassed proprietary counterparts not on general benchmarks, but on a practical, security-critical task. Alibaba and Mistral AI are releasing updates monthly, while Meta has slowed down Llama's development pace – this shifts the race in favor of Chinese and European labs. For companies, this means a real alternative to cloud services: they can deploy a powerful model on their own servers without sending data to the cloud or being constrained by export restrictions.

Sources:

Let's Data Science (letsdatascience.com)
Tech Insider (tech-insider.org)

Sources:

Blog