New open-source and open-weight models like Llama, Qwen and DeepSeek, plus datasets, frameworks and developer tools worth trying.
ByteDance just open-sourced Bernini 2, an omnimodal video pipeline that lets you edit existing footage using natural language, from swapping backgrounds to inserting new characters. It pairs an MLLM planner with a diffusion renderer for fine-grained control, but weighing in at 180GB for the model alone, running it locally will be a heavy lift.
DeepSeek just open-sourced DeepSeek Harness (dsh), an agent framework built entirely on an "everything-is-a-plugin" architecture. Powered by Cordis, it turns models, tools, and sandboxes into modular extensions and spins up an out-of-the-box local Web UI. It is currently in active developer preview, so expect breaking changes ahead.
Xiaomi researchers just open-sourced MiDashengLM-Gen, an end-to-end framework capable of generating coherent audio scenes blending speech, background music, sound effects, and ambient noise in a single pass. Powered by a fine-tuned Qwen3-1.7B and Flow Matching DiT, it turns structured text prompts into unified 16 kHz audio under an Apache 2.0 license.
AlayaLab just open-sourced Evoke, a 14B interactive world model that turns a single image and joystick inputs into real-time, steerable video in just three steps. It maintains persistent scene memory across hour-long sessions and allows mid-flight text prompting. The 57GB model weights and code are publicly available under Apache 2.0.
NVIDIA just launched Nemotron 3.5 Lightning, an open 30B MoE model with only 3B active parameters tailored for high-volume agent tasks like tool calling and result validation. It completes 10,000 tasks 30% faster than Qwen3.6 35B on PinchBench. Frontier models handle planning, while Lightning handles the grunt work.
LTX just launched LTX-2.5, an open-weights video model that generates a 10-second 720p clip in just 6.8 seconds when self-hosted on two Nvidia GB200 superchips. It arrives with native ComfyUI support, native multishot consistency, and a Gemma 4 backbone. It is free for orgs under $10M ARR, giving open-source video a massive speed upgrade.
HKUDS just open-sourced AutoSwarm, a lightweight tool that turns fragmented agents like Claude Code, Codex, and Cursor into one coordinated team. It shares context and memory across tools, delegates parallel tasks based on capabilities, and synthesizes answers into a single output. It works across Windows, macOS, and Linux to end manual agent juggling.
Cactus just rolled out Needle 2, an ultra-tiny 45M-parameter model that packs into a 14MB binary and uses only 28MB of RAM. Built for micro-devices like the Raspberry Pi 5 and sub-$200 phones, it clocks hundreds of tokens per second for local tool calling without needing a GPU.
IndexTeam just released IndexTTS-2.5, upgrading its open-source zero-shot speech synthesis model. The new release adds support for Chinese, English, Japanese, Spanish, and Arabic, featuring finer emotion manipulation, precise phoneme pronunciation overrides, and vLLM production serving.
Meta just released Muse Glimmer, a 30B open agentic model under an Apache 2.0 license. Quantized to roughly 4-bit precision, it shrinks under 20GB to run smoothly on a single consumer GPU or Mac while handling complex tool calls, visual context, and multi-step reasoning. Basically, robust always-on AI agents can finally run entirely offline on local hardware.
Tencent just open-sourced Octop, a self-hosted platform that runs a multi-agent environment from a single local process. It connects directly with chat apps like WeChat, Feishu, and Discord, letting specialized agents collaborate on complex workflows while keeping all conversation history, knowledge bases, and memory strictly on your own machine.
Google DeepMind just open-sourced its WeatherNext models following a Nature paper demonstrating a major breakthrough in cyclone tracking and intensity forecasting. Its three-day forecast matches prior two-day benchmarks—roughly a decade of meteorological progress—and runs 15-day simulations in under a minute on a single TPU.
Multi-Agent CAD (MAC) v2 turns natural-language prompts into editable, printable 3D engineering parts and assemblies. By decoupling planning, coding, and visual verification across multiple agents, it cuts token usage by up to 116x and costs by 13x over baselines. The open-source tool exports STEP, STL, and simulation-ready URDF files.
Prime Intellect just open-sourced Prime Agent, an autonomous coding and research harness designed to continuously improve itself. Built on a recursive language model architecture, it treats prompt context as variables and spawns subagents programmatically inside a persistent Python REPL. Why it matters: It turns ephemeral chat sessions into durable, background-running workflows that refine their own memory and skills over time.
JD just open-sourced JoyAI-Video-Edit, an autoregressive diffusion model that edits video streams on the fly via text instructions. Powered by a 16B-parameter MMDiT, it handles clothing swaps, object removal, and style edits at 30 FPS in 720p, and can run at 24 FPS on a single RTX 5090.
NVIDIA just made its Alpamayo 2 Super autonomous vehicle model available for commercial use under the OpenMDW-1.1 license. Built on Cosmos 3 Super Reasoner with three times the scale of its 10-billion-parameter predecessor, it beats GPT-4o by 23.2 points on LingoQA. It fuses 360-degree camera views to generate driving paths and causal reasoning traces, giving automakers full control over deployment.
Genspark just open-sourced GenOffice, a completely free and ad-free office suite for PC and Mac covering Docs, Sheets, Slides, and PDF editing. Remarkably built by a single engineer in one week using $10,000 in API tokens, the Alpha integrates Genspark's Super Agent to automate research, data crunching, and slide generation inside your workspace.
LongHorizon-Harness introduces a Manage-Execute-Audit loop to prevent agent compounding errors on marathon tasks. Fresh executors handle single subtasks in clean contexts while an independent auditor verifies actual environment changes before logging progress. Swapping in the harness boosted Qwen 3.7-Plus pass rates on WeaveBench from 51.8% to 80.7%.
endless-frontier just launched BigBang-v1 on Hugging Face, an experimental 35B model built on an adversarial synthetic data loop. Instead of relying on human experts to write training tasks, generator and critic agents iteratively designed and solved roughly 10,000 frontier scientific challenges. Why it matters: it rivals trillion-parameter giants on research benchmarks using remarkably little data.
DeepSeek just officially released DeepSeek-V4-Flash-0731 with built-in DSpark speculative decoding. Thanks to heavy post-training, this lightweight model leapfrogs the larger V4-Pro preview across terminal and coding benchmarks, punching well into Opus 4.8 territory while fully open-sourced under an MIT license.
