New open-source and open-weight models like Llama, Qwen and DeepSeek, plus datasets, frameworks and developer tools worth trying.
Meta's FAIR team just released five new AI research models and tools to the public. Highlights include Chameleon for mixed text-and-image generation, JASCO for chord- and beat-controlled music synthesis, and AudioSeal, an audio watermarking system that detects synthetic speech up to 485 times faster.
Apple just released 20 new Core ML models and four datasets on Hugging Face, including FastViT and DepthAnything. Built specifically to run on-device without an internet connection, they let developers build fast computer-vision features while keeping user data strictly private.
DeepSeek just open-sourced DeepSeek Coder V2, an MoE-based code model supporting 338 languages and a 128K context window. It beat GPT-4 Turbo, Claude 3 Opus, and Gemini 1.5 Pro on major coding benchmarks like HumanEval and MBPP+, marking a rare open-source victory over leading proprietary models. Available under the MIT license in 16B and 236B sizes.
The Qwen team just launched Qwen2, an open-source model family spanning five sizes from 0.5B to 72B parameters. Trained on 27 additional languages, the flagship Qwen2-72B goes toe-to-toe with Llama-3-70B across major benchmarks, while context lengths stretch up to 128K tokens. Except for the 72B tier, all smaller models are now released under the Apache 2.0 license.
Researchers at EPFL and Yale School of Medicine just released Llama-3[8B]-MeditronV1.0, fine-tuned on Meta's Llama 3 in 24 hours to help clinicians in low-resource and humanitarian settings. The open-source model beats all open models in its size on MedQA and MedMCQA, with weights and data openly available to ensure equitable access.
Snowflake just launched Arctic LLM, an enterprise-focused open model licensed under Apache 2.0 for research and commercial use. Using a 480B-parameter MoE architecture that only activates 17B parameters at a time, it cost just $2 million to train and targets SQL generation and corporate coding workloads. Snowflake claims it outperforms rivals like Databricks’ DBRX on database tasks.
NVIDIA just rolled out full-stack optimizations for Meta's brand-new Llama 3 across its data center GPUs, edge devices, and PCs. In early benchmarks, a single H200 GPU churned out roughly 3,000 tokens per second on the 70B model, enough to handle 300 users simultaneously. The 8B model is also landing on RTX PCs and Jetson Orin modules.
Hugging Face just released Idefics2, an open-source 8B vision-language model licensed under Apache 2.0. It preserves native image resolutions, significantly upgrades OCR, and punches way above its weight class on visual QA benchmarks against 30B+ rivals. Why it matters: Developers get an efficient multimodal base model ready to fine-tune directly in Transformers.
Mistral AI just dropped its largest open-source model yet, releasing Mixtral 8x22B via an unceremonious torrent link on X. The sparse MoE model features 176 billion total parameters with 44 billion active per token and a 65,000-token context window. Released under a permissive Apache 2.0 license, it delivers massive compute efficiency to open-source builders without corporate guardrails.
Meta just confirmed that its next-generation open-source model, Llama 3, will begin rolling out within the next month. The company plans to release multiple versions across the year and integrate them throughout its product lineup. Llama 3 aims to relax previous over-cautious restrictions, with reports expecting parameter sizes to reach roughly 140B.
Google just released CodeGemma in collaboration with Hugging Face, bringing a family of open-access code specialists built on Gemma. Trained on an extra 500 billion tokens, the lineup includes an infilling-focused 2B model that runs on consumer hardware alongside 7B base and instruct variants. Why it matters: It offers developers fast, private code completion without needing massive cloud infrastructure.
Stability AI just launched Stable LM 2 12B, a multilingual model trained on 2 trillion tokens across seven European languages. Available in both base and instruction-tuned flavors, it is built to handle function calling and RAG workflows on standard hardware. The compact 1.6B model also got an upgrade with better conversation skills and tool use.
Stability AI just released Stable Video 3D (SV3D), adapting its video diffusion tech to turn a single object image into consistent multi-view orbital videos and 3D meshes. It comes in two flavors—one for generic orbits and another supporting custom camera paths. Weights are open on Hugging Face for non-commercial use, while commercial deployment requires a membership.
xAI just released the base model weights and architecture of its Grok-1 language model under the permissive Apache 2.0 license. This is the raw pre-training checkpoint from October 2023, trained from scratch as a Mixture-of-Experts system without fine-tuning for specific tasks like dialogue. Developers can now access instructions directly on GitHub.
Stability AI and Tripo AI just released TripoSR, an open-source model that transforms a single 2D image into a textured 3D mesh in under half a second. It runs fast even without a dedicated GPU, and both the code and weights are completely free for commercial use under an MIT license.
Google just released Gemma, a new lightweight family of open models built on the same research as Gemini. Available in 2B and 7B parameter sizes, it runs directly on laptops or desktops, beats rival open models at its scale, and comes with permissive commercial licensing. It gives local and open-source AI developers a massive new alternative to Llama.
Apple just teamed up with UCSB researchers to release MGIE, an open-source model that lets you edit photos by simply typing out what you want. Instead of tweaking sliders, you can give prompt-based commands like "make it more healthy" to add veggies to a pizza, or ask it to brighten specific dark spots. It is already available on GitHub and Hugging Face Spaces.
Abacus AI just released Smaug-72B, a fine-tuned version of Alibaba's Qwen-72B that is turning heads across the open-source community. It is the first open-source model to score an average above 80 on Hugging Face's Open LLM leaderboard, outpacing proprietary heavyweights like GPT-3.5 and Mistral Medium in key reasoning benchmarks. The weights are freely available now on Hugging Face.
Mistral CEO Arthur Mensch just confirmed that the mystery open-source model “miqu-1-70b”—which scored near GPT-4 on benchmarks after surfacing on 4chan and Hugging Face—is indeed an old Mistral checkpoint retrained from Llama 2, leaked by an early customer’s employee. Why it matters: it proves open-source AI is closing the frontier performance gap much faster than expected.
Apple and Cornell researchers quietly open-sourced Ferret, an MLLM that can analyze and answer queries about any freeform region inside an image. Users can highlight specific elements to trigger detailed contextual responses. It marks a rare open-source push from the usually secretive tech giant, though it is currently restricted to non-commercial research.
