New open-source and open-weight models like Llama, Qwen and DeepSeek, plus datasets, frameworks and developer tools worth trying.
Hugging Face just launched SmolVLM2, a family of lightweight multimodal models coming in 2.2B, 500M, and 256M parameters. The 500M version delivers video understanding comparable to the 2.2B model at a fraction of the memory footprint. With day-zero MLX and Swift support, these models can run completely offline on an iPhone or inside a free Google Colab.
Perplexity just open-sourced R1 1776, a post-trained variant of DeepSeek-R1 designed to remove political censorship without sacrificing performance. Trained with Nvidia NeMo on a curated 40,000-prompt dataset, it delivers factual answers across sensitive topics while keeping original math and reasoning strengths intact.
Baidu just announced plans to open-source its upcoming Ernie 4.5 large language model series on June 30. The models will roll out to users over the coming months before the code is released. The shift comes shortly after DeepSeek-R1 shook up the industry, putting pressure on tech giants to rethink their open-source AI strategies.
Tencent just open-sourced Hunyuan3D 2.0, a two-stage foundation system for generating detailed, textured 3D meshes from single images. By decoupling geometry creation via a DiT shape generator from texture baking with a dedicated diffusion model, it can also texture handcrafted meshes, beating previous open and closed-source baselines.
MiniMax just open-sourced its new MiniMax-01 series, featuring a 456B-parameter language model and a visual multimodal model. Built on a novel Lightning Attention architecture with linear complexity, the models handle up to 4 million tokens of context with virtually zero loss. Weights are available on GitHub alongside access via MiniMax Chat.
UC Berkeley’s NovaSky team just open-sourced Sky-T1-32B-Preview, a reasoning model that rivals OpenAI's o1-preview while costing under $450 to train. Fine-tuning Qwen-2.5-32B on 17K curated reasoning samples took 19 hours on 8 H100 GPUs. Best of all, the entire pipeline, data, and weights are publicly available.
Adobe Research and HKUST just open-sourced TransPixar, a new model that solves a major VFX headache: generating videos with built-in alpha channels. Instead of solid backgrounds, it generates transparent RGBA elements like smoke and portals ready for compositing. It is an early-stage research project, but the code and interactive demo are already live.
Microsoft just released the official weights for its 14B-parameter Phi-4 model on Hugging Face under a permissive MIT license. Previously restricted to the Azure AI Foundry platform, developers can now freely download, fine-tune, and deploy the high-reasoning model locally for commercial applications.
Researchers at the University of Surrey just open-sourced NitroFusion, an AI model that generates images in near real time as you type using standard consumer hardware. The framework lets users dial between one and four refinement steps, balancing speed and image quality without cloud GPUs or monthly subscription fees.
Meta just introduced Llama 3.3 70B, a lightweight text model that delivers the punch of its massive 405B predecessor while running significantly cheaper. Armed with advanced post-training, it tops GPT-4o and Gemini 1.5 Pro on key benchmarks like MMLU, and is available for download right away.
Google just launched PaliGemma 2, its next-gen vision-language model family built on Gemma 2. It offers drop-in replacement upgrades across multiple model sizes without major code changes, making it significantly easier for developers to fine-tune visual AI for custom datasets and vision tasks.
Andrew Ng’s team just released aisuite, an open-source Python library that lets developers switch between models from OpenAI, Anthropic, Ollama, and more by changing a single string. It wraps fragmented APIs into one standardized interface, drastically cutting the friction of building multi-model apps.
Anthropic just open-sourced Model Context Protocol (MCP), a new standard designed to hook any AI model directly into tools, code repos, and enterprise software. Instead of custom connectors for every app, developers can build two-way connections to services like GitHub and Slack in under an hour. The protocol is open to rivals, but whether competitors like OpenAI adopt it remains an open question.
Lightricks just released LTX Video, an open-source model that generates five seconds of video in just four seconds. With only 2 billion parameters, it runs near real-time on consumer GPUs like the RTX 4090 instead of locking users behind paid APIs. Why it matters: Bringing rapid video generation to local gaming PCs gives indie creators a real alternative to big tech monopolies.
Tether just launched an open-source Wallet Development Kit (WDK) that lets developers embed non-custodial USDT and Bitcoin wallets into any app or device. Notably, it is explicitly built to support autonomous AI agents and robots alongside humans, giving software systems full control of their own crypto without third-party custody.
The Qwen team just released the full Qwen2.5-Coder lineup, spanning six sizes from 0.5B to 32B parameters. The flagship 32B-Instruct model matches GPT-4o on major coding benchmarks like EvalPlus and Aider, supporting 40+ languages and 128K context. Most models run on an Apache 2.0 license, bringing serious frontier-level coding power directly to local hardware and Cursor setups.
Microsoft just open-sourced Magentic-One, a multi-agent system built on AutoGen where a lead Orchestrator directs four specialized agents to browse the web, read files, code, and execute terminal commands. It delivers top-tier performance on benchmarks like GAIA without bespoke architectures. The modular design lets developers easily swap or add new agents.
The Open Source Initiative just launched version 1.0 of its Open Source AI Definition. To qualify, models must share complete training code and detailed data provenance without restrictive commercial caps, effectively disqualifying Meta's Llama. Why it matters: regulators and developers now have a strict standard to call out open-washing.
Meta just dropped NotebookLlama, an open-source project that recreates Google NotebookLM's viral podcast generator using its own Llama models. The pipeline turns uploaded files into dramatic dialogue scripts before piping them into open text-to-speech tools. Why it matters: The audio still sounds robotic and cuts each other off, but developers now have an open recipe to customize and upgrade.
Nvidia just released the NVLM 1.0 family of multimodal models, making weights public for its 72B flagship NVLM-D-72B. The model matches frontier proprietary systems like GPT-4o across vision tasks and unexpectedly gained 4.3 points in text benchmarks post-multimodal training, giving open-source developers frontier-grade power.
