Open-source LLMs like GLM-5.2, Qwen, and LongCat gain traction, challenging proprietary giants and enabling local deployment. China promotes open models as a public good, while Hugging Face reports nearly half of Fortune 500 companies use open-source AI.
Hugging Face CEO Clem Delangue highlights the golden moment of open source AI, now used by roughly half of Fortune 500 companies. An opportunity to reflect on open source's role in the enterprise ecosystem, especially for those seeking control, data ...
The freely downloadable GLM-5.2 sparks alarmist headlines for its cybersecurity prowess. But the real story is the shift toward LLMs that run on commodity hardware, without middlemen, redefining the balance of power, privacy, and accountability for t...
At the UN’s first global dialogue on AI governance, China framed open source models like DeepSeek and Qwen as a shared asset that lowers adoption barriers. This stance directly impacts data sovereignty and on-premise deployment, reshaping TCO for tho...
Chinese startup MiniMax is readying its M3 Pro model, a 2.7-trillion-parameter open-source LLM expected as soon as Q3 2025. With a focus on complex reasoning, it dwarfs the 428-billion-parameter M3 and reignites the debate on the accessibility of fro...
Baidu releases Unlimited-OCR on ModelScope: 3.3 billion parameters, MIT license, one-shot parsing of images, PDFs, and multi-page documents. 32K output length, Transformers inference and SGLang serving with OpenAI-compatible streaming. A building blo...
Microsoft released FastContext, a 4B-parameter subagent for repository exploration in LLM coding workflows. It cuts token usage by up to 60%, boosts SWE-bench accuracy, and can now run on-prem via a pull request for 'oh my pi'. A signal for those eva...
AllenAI unveils TMax, an open dataset of RL environments and a training recipe that yields compact terminal agents up to 27B parameters. The 9B model beats all open sub-10B contenders on Terminal Bench 2.0 and approaches closed systems like Claude Ha...
After firing AI chief Junyang Lin, Qwen has kept its 3.7 model line fully closed source, making it the only major Chinese lab without a recent open release. As competitors like DeepSeek and GLM continue shipping, the move narrows choices for on-premi...
Noema AI has released Noema Atlas, an open-source peer-to-peer software for LLM weight distribution. Built on Iroh and BLAKE3-based content-addressing, it enables verified machine-to-machine transfers, automatic deduplication, and optional fallback t...
Ohio State University released QUEST-35B, an autonomous research agent trained on 32 H100 GPUs and synthetic data. Code, weights, and training recipe are public, with competitive benchmarks against closed systems. A signal for on-premise deployment.
The 2-bit quantized GLM-5.2 shrinks from 1.51TB to 238GB while retaining ~82% accuracy. It can now run locally on a 256GB Mac or systems with enough RAM/VRAM via llama.cpp and Unsloth Studio, opening new possibilities for on-premise AI deployment.
Hugging Face is offering free inference for the GLM-5.2 model for the next six hours. This limited-time initiative highlights the dynamics of Large Language Model deployment and cost considerations. For companies evaluating on-premise solutions, mana...
GLM-5.2 has been recognized as the top "open weight" Large Language Model (LLM) for creative writing, according to Sam Paech's benchmark on EQ Bench. This achievement highlights the potential of accessible models for on-premise deployment scenarios, ...
Z.ai has open-sourced its GLM 5.2 model, generating significant community excitement. Developers and enterprises are now eagerly anticipating a "Flash" series successor, ideally within the 27 to 120 billion parameter range, to optimize on-premise and...
The release of GLM 5.2, a 744-billion-parameter Large Language Model under an MIT license, marks a significant development for on-premise AI. While the full model necessitates enterprise-grade clusters, its potential for distillation and fine-tuning ...
