A wave of powerful open-source LLMs with permissive licenses and advanced quantization methods is democratizing AI, enabling cost-effective and sovereign deployment on local hardware.
Meituan has open-sourced LongCat-2.0, a fresh component of China’s homegrown AI stack. The release points to a maturing parallel ecosystem focused on data sovereignty and on-premise deployments, lessening reliance on US cloud vendors.
The rise of open source models isn't eroding the position of frontier labs like Anthropic, because each occupies a different phase of enterprise adoption. An analysis of the structural incentives reshaping AI deployment.
NVIDIA Nemotron Labs released Puzzle-75B-A9B, a deployment-optimized hybrid MoE LLM that cuts one-third of the parent model's total parameters while preserving accuracy and doubling throughput on 8×B200. On a single H100, 1M-token context concurrency...
NVIDIA releases Audex-30B-A3B, a mixture-of-experts LLM with 30 billion parameters that extends Nemotron-Cascade-2 to audio without regressing on text capabilities. It handles recognition, translation, text-to-speech, and audio generation, with up to...
Sberbank’s new Mixture of Experts model – 432B total, 28B active per token – debuts with a GGUF edition ready for consumer hardware. A strong signal for teams building on-premise stacks around large language models.
Tencent has open-sourced the Hy3 model: 295 billion total parameters, 21 billion active, mixture-of-experts architecture, and — crucially — an Apache 2.0 license. It replaces the former community license that prohibited use in South Korea, the UK, an...
The weights of longcat 2.0 (1.6 trillion total parameters, 48 billion active) are now available under the MIT license. A strong signal for teams evaluating on-premise LLM deployments, with implications for data sovereignty, customization, and TCO.
The French startup has raised significant funding since 2023 to bring frontier models to the public. But the real differentiator is its open models: a choice that redefines data control and deployment strategies for enterprises.
The Portuguese government has released Amalia, a 9 billion-parameter Large Language Model under an Apache 2.0 license. Available in SFT and DPO variants, it aims to strengthen the country's digital sovereignty. The lack of coding benchmarks doesn't h...
Mistral has released Leanstral 1.5, an Apache 2.0-licensed LLM focused on formal verification. With only 6 billion active parameters out of 119 billion total, it achieves state-of-the-art results on miniF2F, PutnamBench and FATE, and discovered 5 bug...
An Nvidia AI pioneer dismisses AGI and likens OpenAI and Anthropic's closed models to AOL and Prodigy's walled gardens. The bet is on open, customized LLMs, with deep implications for those managing sensitive data on-premises.
Z.ai has released GLM-5.2, ranking fourth in performance benchmarks, with coding and agentic capabilities close to market leaders. Its cost is a fraction of Anthropic or OpenAI, raising questions about how this will influence deployment choices, espe...
The new SenseNova-U1-8b-MoT-Infographic-V2 excels at generating and editing dense infographics. Released under Apache 2.0, it outshines its only rival, Ideogram 4, thanks to deployment freedom. It requires up to 36 GB VRAM, but quantized versions dro...
With just over 10 million people, Portugal has released Amália, an LLM trained for European Portuguese. The government open-sources the model and its training data, betting on linguistic and digital sovereignty — a move that reignites the debate on d...
A 92B-total, 6B-activated MoE model with a 512k-token context window, trained on Ascend hardware. It features hybrid DSA/SWA attention, multi-token prediction, and the Muon optimizer to reduce on-premise inference costs.
