AI News Brief

1. Meituan Open-Sources LongCat-2.0: 1.6T-Parameter Coding Model Trained Entirely on Chinese Chips

Meituan released LongCat-2.0, a 1.6-trillion-parameter Mixture-of-Experts agentic coding model under the MIT license, with a 1 million token context window. The model was trained entirely on a cluster of over 50,000 domestic Chinese ASICs (no Nvidia, no AMD), proving near-frontier models can scale without US chips. It scored 59.5 on SWE-bench Pro and 70.8 on Terminal-Bench, and has been leading the OpenRouter usage leaderboard.

2. Zhipu's GLM-5.2 Matches Claude Mythos on Cybersecurity Tasks at a Fraction of the Cost

Zhipu AI's open-weight GLM-5.2 model is performing on par with Anthropic's restricted Claude Mythos on cybersecurity and vulnerability detection tasks. It trails Anthropic's Opus 4.8 by just one percentage point on a key agentic benchmark while costing roughly a fifth as much to run. GLM-5.2 leads open-weight models on SWE-bench Pro at 62.1 and ranks among the top 10 most-used models on OpenRouter. Semgrep benchmarks found GLM-5.2 actually outperformed Claude Code on IDOR vulnerability detection (higher F1 score).

3. DeepSeek Releases DSpark Open-Source Inference Framework, Slashing Compute Costs by Up to 85%

DeepSeek released DSpark, an open-source inference optimization framework that accelerates AI model generation speeds by up to 85% without requiring hardware upgrades. The framework is designed to cut inference compute costs significantly, making AI deployment more accessible and affordable for developers running open-weight models on existing infrastructure.

4. Trump Administration Forces OpenAI to Restrict GPT-5.6 Release, Creating De Facto Model Licensing

OpenAI limited the release of GPT-5.6 Sol to roughly 20 government-approved partners at the request of the Trump administration, which is conducting cybersecurity reviews under a new executive order. This follows a similar directive that forced Anthropic to suspend access to its Fable 5 and Mythos 5 models. The de facto licensing regime has no publicly defined safety standards and has split Trump's pro-AI allies, with former White House AI adviser David Sacks warning the US is "deviating" from its pro-innovation strategy. Chinese open-source model usage has surged on OpenRouter as a direct result.

5. Palantir and NVIDIA Deploy Nemotron Open Models Inside US Government Systems

Palantir launched an intelligent engine for running NVIDIA Nemotron open-source models in sovereign, air-gapped government environments, giving US agencies full control over model weights, training data, and deployment. The initiative is already in use by multiple US clients supporting critical infrastructure. NVIDIA CEO Jensen Huang framed open-source AI as "foundational to national security," and Palantir CEO Alex Karp said it removes security risks of proprietary insights migrating into closed models. Palantir stock rose 3.85% on the announcement.

6. South Korea Announces $880 Billion AI Chip Investment Plan Led by Samsung and SK Hynix

South Korea unveiled a massive national investment plan totaling at least 1,350 trillion won ($880 billion) over 10 years, with Samsung Electronics and SK Hynix committing to build a new chipmaking hub. President Lee Jae-myung framed it as a "national survival" strategy for the AI era, targeting semiconductor fabs, data centers, and AI infrastructure hubs outside Seoul. The plan is one of the largest single-country AI infrastructure commitments ever announced.