Unsloth enables training of OpenAI's gpt-oss with RL and GRPO, achieving 3x faster inference, 50% less VRAM, and 8x longer context without accuracy loss.
SimpleFold: Folding proteins is simpler than you think
SimpleFold is a groundbreaking protein folding model utilizing general purpose transformer layers and achieving 3B parameters, trained on over 8.6M protein structures, marking it as the largest folding model to date.
Moondream 3 Preview: Frontier-level reasoning at a blazing speed
Moondream 3 introduces a 9B MoE architecture with 2B active parameters, achieving frontier-level visual reasoning while maintaining fast and efficient inference for real-world applications.
First Malicious MCP in the Wild: The Postmark Backdoor Stealing Your Emails
The postmark-mcp package, downloaded 1,500 times weekly, has been found to exfiltrate emails to an external server due to a malicious line of code added in version 1.0.16, highlighting vulnerabilities in trusting third-party tools.
Modular Manifolds
Normalization of tensors is crucial in training large neural networks to prevent issues like numerical underflow and overflow, with techniques such as layer norm and the Muon optimizer providing effective solutions for maintaining tensor health.
Why We Think
Test-time compute significantly enhances model performance by allowing longer reasoning periods, akin to human cognitive processes, as demonstrated in recent studies on Chain-of-Thought (CoT) prompting and reinforcement learning techniques.
[R] DynaMix: First dynamical systems foundation model enabling zero-shot forecasting of long-term statistics at #NeurIPS2025
The DynaMix model, accepted at #NeurIPS2025, enables zero-shot forecasting of long-term statistics from minimal context, outperforming traditional time series models with only 0.1% of the parameters and over 100x faster inference.
Windows ML is generally available
Windows ML is now generally available, enabling developers to deploy local AI solutions efficiently across a wide range of Windows devices, leveraging the latest advancements in silicon and software integration.
LLM Observability in the Wild – Why OpenTelemetry Should Be the Standard
LLM observability is hindered by competing standards, with OpenTelemetry being the industry standard but lacking AI-specific insights, while OpenInference offers richer span types but struggles with compatibility and language support.
[P] Why MissForest Fails in Prediction Tasks: A Key Limitation You Need to Keep in Mind
The MissForest algorithm fails in predictive tasks due to its inability to save imputation models, resulting in data leakage during train/test splits, which undermines model integrity.