Strong debut for Llama 3.1 405B! Tying for 3rd overall on the lmsys.org leaderboard today! Tied for first place in English-only evals, coding and a few others 🏆 Wins for 70B & 8B as well, #9 and #29 respectively, both punching well above their weights (pun intended).

Excited for this!

We’re opening a call for applications for the next round of Llama Impact Grants. The program supports organizations pursuing ideas to use Llama to address social challenges in their communities with $2M in grants available. Full details + criteria: ai.meta.com/blog/llama-3-1…

Fascinating, how entire LLM industry is chasing ELO score on lmsys, just recently it was Open LLM leaderboard and MMLU, and still around those who remember the days of GLUE and SuperGLUE. Meanwhile Goodhart's law never gets old: "When a measure becomes a target, it ceases to be

As AI models become increasingly complex we’ll need new benchmarks to assess more advanced behaviors. I'm excited to help push this work forward with this new research grant program from our team!

Really cool new work from Meta. SAM 2 is the first unified model for real-time, promptable object segmentation in images & videos — and it’s Apache open source so that anyone can build on it. Congrats Nikhila Ravi + the whole team! sam2.metademolab.com

Cool new work from our friends at NVIDIA AI, they just released Llama-Minitron 3.1 4B and shared a detailed breakdown of how they built it by pruning and distilling Llama 3.1 8B. Exciting work Bryan Catanzaro! developer.nvidia.com/blog/how-to-pr…

Congrats Snowflake team on the preview launch of Cortex Analyst! Thrilled to see these use cases that reiterate the value and industry wide impact of open source.

Pretty cool work on tuning-free personalization that enabled "imagine me" in Meta AI

If you’re building with AI in Europe, you know just how important this discussion is right now. Open source AI creates more opportunities for all and we hope that the EU doesn’t limit the possibilities for new innovation for European consumers and businesses.

A good read from the Cursor team: cursor.com/blog/instant-a… A few notable points: - They created a synthetic dataset with open source models. - They trained Deepseek Coder and Llama 3 families of models. - Fine-tuning Llama 3, they’re reaching coding agent performance nearly

It’s awesome to watch the pace of innovation from our partners in the Llama ecosystem. Cerebras just launched a high-speed Inference service for Llama 3.1 that delivers 450 tokens/s for 70B and 1,850 tokens/s for 8B. Developers are going to have a lot of fun with this.

New work from @NVIDIA, NIM Agent Blueprints will be a big unlock for the next wave of enterprises taking control of their own AI stack with Llama 405B. nvidianews.nvidia.com/news/nvidia-an…

Llama is fast becoming the industry standard for open source AI w/ nearly 350M downloads to date & usage more than doubling in the last four months across our cloud partners. Excited about what’s ahead and thankful to the community building with us! ai.meta.com/blog/llama-usa…

Good read for anyone else who starts their morning checking the lmsys.org leaderboard. Tl;dr: controlling for the effect of length and style in the leaderboard shows some really interesting dynamics from substance — a single metric isn’t enough to tell the whole story.

On the heels of the update I shared on Llama yesterday, we’re also seeing Meta AI usage growing FAST with 185M weekly actives! 🚀

This is the magic of Llama and open source.