NVIDIA Debuts DLSS 5 as Mistral AI Unveils Small 4 and Leanstral Models
Large Language Models & Releases
Mistral AI Releases Mistral Small 4 and Teases Mistral 4 Family: Mistral AI has launched Mistral Small 4, a 119B parameter model utilizing a Mixture of Experts (MoE) architecture with 6.5B active parameters per token. The model is multimodal, supports a 256k context length, and is released under the Apache 2.0 license, while rumors and GitHub leaks suggest a broader "Mistral 4" family focusing on reasoning and development capabilities.
- Mistral 4 rumors
- Mistral Small 4:119B-2603
- Mistral 4 Family Spotted
- Introducing Mistral Small 4
- Mistral 4 spotted on GitHub
Leanstral: An Open-Source Code Agent for Lean 4: Mistral AI has released Leanstral, the first open-source code agent designed specifically for the Lean 4 proof assistant. The model is optimized for expressing complex mathematical objects and software specifications, aiming to advance verifiable "vibe-coding" and open research.
- [Model Release] Leanstral](https://www.reddit.com/r/MistralAI/comments/1rvkkkz/model_release_leanstral/)
- Leanstral.
NVIDIA Nemotron 3 Benchmarks and Performance: NVIDIA's Nemotron 3 Ultra Base model has shown superior throughput and reasoning accuracy in benchmarks against competitors like Kimi K2 and GLM. However, community testing of the smaller 4B variant indicates it may struggle with reasoning and code generation compared to rivals like Qwen 3.5.
- INCREDIBLE STUFF INCOMING
- I was hyped for Nemotron 3 4B and it completely disappointed me compared to Qwen 3.5 4B
Covenant-72B: Largest Model Trained on Decentralized Nodes: Covenant AI has released Covenant-72B, a new model notable for being the largest to date trained on decentralized, permissionless GPU nodes. The training process utilized the SparseLoco method to efficiently manage communication overhead across the distributed network.
- 1Covenant/Covenant-72B: Largest model so far to be trained on decentralized permissionless GPU nodes
Partnerships & Industry News
Mistral AI and NVIDIA Announce Strategic Partnership: Mistral AI and NVIDIA have partnered to co-develop frontier open-source AI models. This collaboration is intended to accelerate the development of advanced AI technologies by leveraging NVIDIA's hardware and Mistral AI's model expertise.
Graphics & Gaming AI
NVIDIA DLSS 5 Introduces AI-Generated Visual Fidelity: NVIDIA has announced DLSS 5, a breakthrough technology that uses AI models to generate photorealistic lighting and materials in real-time. This version provides game developers with detailed controls to maintain specific visual aesthetics while significantly enhancing overall graphical quality.
AI Research & Community Insights
Researcher Identifies Universal "Danger Zone" in Transformer Architectures: A community experiment involving layer surgery on six different model architectures discovered a "danger zone" at roughly 50-56% depth that causes models to fail. The findings provide practical guidelines for local developers regarding optimal layer duplication and the architectural limits of deep learning models.