a printed flyer on a pole in new york city that reads “Open Source”. The bottom is cut into small strips that have the word “Mistral” “Hugging Face” “Gemma” “Llama” and “Phi” ideogram.ai
Databricks
“Some personal news: I’m thrilled to have joined @Databricks @DbrxMosaicAI as a Research Scientist last month, before I start as MIT faculty in July 2025! Expect increased investment into the open-source DSPy community, new research, & strong emphasis on production concerns 🧵.” / X
Gemma
“Seems GrokAdamW is an improvement. This is training gemma-2-2b with Dolphin 2.9.4 dataset, all settings equal
Grok
“I got early (beta) access to Grok 2. Spent all night testing it. Here are 6 things it can do that ChatGPT cannot:
“xAI began rolling out early beta access for Grok 2, the new AI model that leverages real-time data from X I tested it and found 6 core features that it can do that ChatGPT can’t (in the thread below)
Hugging Face
“HuggingFace ships!! From hardware DIY manual to Jupyter notebooks, you get the whole deluxe tutorial experience. Everything that moves will eventually be autonomous. We need a new generation of talents to work on bridging the world of bits with the world of atoms, and open-source” / X
Meta/Llama
“Meta changed the game with the release of Llama 3.1 405B! Now, you can easily deploy @AIatMeta Llama 3.1 405B on @googlecloud Vertex AI! 🚀 This means you GPT-4 level capabilities running in-house, giving you full control. ✅ TL;DR: 🧠 Meta Llama 3.1 405B offers 128K token
“Using Medusa, we achieved a 94% to 122% increase in tokens per second for Llama 3! 🤯 Medusa is a method for generating multiple tokens per forward pass during LLM inference. After trying more fundamental optimizations (like quantization, using H100 GPUs, or TensorRT-LLM), more
⛈️ Llama-3.1 Storm Models – a akjindal53244 Collection
Mistral
NVIDIA Releases Small Language Model With State-of-the-Art Accuracy | NVIDIA Blog
“Nvidia and Mistral just released Mistral-NeMo-Minitron 8B, a small language model that can run on laptops and PCs It outperforms Mistral-7B and Meta-LLama 3.1-8B on the Open LLM leaderboard Small models are improving at an insane rate 👀
“👀Mistral-NeMo-Minitron 8B, one of the most advanced open models in its size, delivers higher accuracy and lower computational cost, with leading performance on nine benchmarks. Deep dive to get started ➡️
[2408.11796] LLM Pruning and Distillation in Practice: The Minitron Approach
“Microsoft just 🎤🫳 3 new models – Phi 3.5 mini instruct (3.8B, 128k context length) – Phi 3.5 MoE (42B-A6.6B, 128k context) – Phi 3.5 Vision instruct (VLM) All MIT Licensed. Let’s go 🚀
“Should we prune & distill big models into smaller models, or should we train smaller models? 🤔 The release of Mistral Nemo Minitron 8B might show that we might want to the first one. @ailozovskaya evaluated the model on the Hugging Face Open LLM Leaderboard benchmarks,
Phi
“How the hell Phi-3.5 is even possible? Phi-3.5-3.8B (Mini) somehow beats LLaMA-3.1-8B.. (trained only on 3.4T tokens) Phi-3.5-16×3.8B (MoE) somehow beats Gemini-Flash (trained only on 4.9T tokens) Phi-3.5-V-4.2B (Vision) somehow beats GPT-4o (trained on 500B tokens) how? lol
microsoft/Phi-3.5-mini-instruct · Hugging Face
Phi-3CookBook/md/01.Introduce/Phi3Family.md at c53fa9fda5df6a42476dd8ba5f1ccb446dd1608c · microsoft/Phi-3CookBook · GitHub
Other Open Source News
“🎉 Based in Austin and want to meet other LangChain enthusiasts? 🤓 Join the community-run @AustinLangChain User Group this Wednesday, August 21st at 6 PM for their Hacky Hour. This casual gathering is the perfect opportunity for AI and LangChain enthusiasts to come together,
“📚 LangChain JS Routing with LLM Classifiers Check out this hands-on tutorial on using LangChain JS to dynamically select prompts based on the query type. 🥕carrots, 🍝lasagna, or general questions—all handled seamlessly with LLM classifiers!
“Organize your Workspace in LangSmith with resource tags. 🏷️ Resource tags help you efficiently manage, group, search, and filter through resources in your workspace. A resource can be a project, dataset, prompt, annotation queue, or deployment. You can now label resources
“You can now distill pretrained Transformers to Mamba / hybrid architecture to get really strong models with fast inference in just a few billion tokens. Beautiful math as always” / X
“AI21 released 🥁 Jamba 1.5 – Mini (52B – 12B active) + Large (398B – 94B active) – Instruct models – Long context (256k) – Strong quality – Multilingual – Function call, JSON output, document understanding – Hybrid transformer/Mamba arch 🔥Read more at
“@giffmana @ahatamiz1 @ylecun mambaout_tiny – 82.7% – 3800 (2132) im/sec vit_little_patch16_reg4_gap_256 – 82.5% – 3246 (2722) im/sec mamba_vision_T2 – 82.7% – 2660 im/sec Put mambaout in there because it’s simple and seems to work pretty well. There are 3090 numbers, don’t have a free A100/H100 to bm on.” / X
“AI21Labs just released Jamba 1.5 Mini & Large – MoE, permissively licensed, 256K context, Multilingual, JSON model & Tool use 🔥 > Jamba 1.5 Large – (94B active, 398B total) > Jamba 1.5 Mini – (12B active/ 52B total) > Arena Hard: Large (65.4), Mini (46.1) > MMLU: large (81.2),
[AINews] Jamba: Mixture of Architectures dethrones Mixtral • Buttondown
“Sometimes you just want a RAG stack that works. BeyondLLM by @aiplanethub has nice abstractions on top of @llama_index that lets you build an advanced RAG pipeline with full evaluation, observability, and advanced RAG features in 5-7 lines of code! Advanced RAG features: –
“LlamaIndex 0.11 is launched! We’ve added literally hundreds of features and bug fixes since 0.10 back in February, and are continuing our push to make LlamaIndex the production-ready platform you want. Headline features: ⭐️ Workflows replace Query Pipelines! ⭐️ In 0.11, our
“Cool new project: MLX Hub A command-line tool that let’s you search, download, and manage MLX models from the @huggingface Hub. Install: pip install mlx-hub Repo/docs:
“Nous Research released Hermes 3, a new open-source model It’s available in three sizes (8B, 70B, and 405B) with the 405B parameter model achieving SOTA relative to other open models Open source AI is still having its moment
“Fine-tuning models? Here’s what you need to know: 1. Data is the most important factor, not the model itself. Focus on thoroughly preparing and testing your data. 2. Avoid vendor lock-in. Ensure you can always export the model weights to something like huggingface 3. Start” / X
“Announcing Dracarys – The Best Open-Source LLM For Coding (70B Class) Open-source LLMs are bad at coding; Dracarys is our foray into fixing the problem! Most open-source fine-tunes severely deteriorate overall performance while improving the base model on some narrow
“🧮 Qwen2-Math-72B is really strong
meetkai/functionary-medium-v3.1 · Hugging Face
“🚀 Introducing Hugging Face Similar: a Chrome extension to find relevant datasets! ✨ Adds a “Similar Datasets” section to @huggingface dataset pages 🔍 Recommendations based on dataset READMEs 🏗️ Powered by @trychroma and @SnowflakeDB Get it now!
“We released the #Jamba 1.5 open model family: – 256K #contextwindow – Up to 2.5X faster on #longcontext in its size class – Native support for structured JSON output, function calling, digesting doc objects & generating citations
“We released the #Jamba 1.5 open model family: – 256K #contextwindow – Up to 2.5X faster on #longcontext in its size class – Native support for structured JSON output, function calling, digesting doc objects & generating citations
“We wanted to share some more granular details about the Jamba 1.5 model family – and specific benchmarks on latency, context window, and quality. [1/6]” / X
Leave a Reply