Skip to content
IFAI

Open source

New open-source and open-weight models like Llama, Qwen and DeepSeek, plus datasets, frameworks and developer tools worth trying.

14 stories in the last 7 days

  1. Reach 48

    Fudan and Tencent open-source Prism for native 2K video-audio generation

    Researchers from Fudan University, Tencent Hunyuan, and Zhejiang University just open-sourced Prism, a framework designed for native 2K joint video-audio generation. Instead of wasting compute across every pixel, Prism uses dynamic sparse attention to focus on sound-producing and rapidly changing visual regions, speeding up training by 2.5x while keeping audio and video tightly synchronized. Preview weights and inference scripts are already available on Hugging Face.

  2. Reach 51

    Meta wants you to build your own Muse AI gadgets

    Meta just open-sourced its Muse AI SDK, letting developers build custom AI agent hardware using off-the-shelf parts like Raspberry Pi and ESP32 boards. Suggested DIY builds include color E Ink reminder displays and portable touchscreens. Meta is also giving away 5,000 units of its pre-built Muse Home Link smart controller to jumpstart home automation use cases.

  3. Reach 48

    Cactus packs a full speech-to-text model into just 16.9 MB

    Cactus Compute just released Whistle, a tiny speech recognition model that fits into a single 16.9 MB file. It runs entirely on the CPU with zero dependencies, delivers transcription and word-level timestamps across seven languages, and beats the 145 MB Whisper base model on several standard benchmarks.

  4. Reach 49

    Strands drops a 2B model that ditches text to decide in 115ms

    Strands just open-sourced Decider 2B, a small model built strictly for fast decision-making rather than chatting. Built on a Qwen3.5-2B backbone with a custom pointer head, it picks between predefined options with confidence scores in a median 115ms on an RTX 3090. Why it matters: It gives local agents cheap, reliable guardrails before firing expensive LLM calls.

  5. Reach 48

    Terra simulates real muscles and bones for AI locomotion

    Researchers just open-sourced Terra, an AI system that teaches simulated bodies to navigate complex terrain using realistic musculoskeletal mechanics rather than simple joint controls. Trained on over 9 hours of motion data, it walks across slopes and steps, and can be fine-tuned for jumps and rolls under the Apache 2.0 license.

  6. Reach 48

    Nvidia's PixelUMM kills the VAE for raw pixels

    Nvidia and the University of Waterloo just introduced PixelUMM, an open-source model that handles image and video understanding and generation directly in raw pixel space without a VAE or vision encoder. Built on Qwen3-8B, it runs all four multimodal tasks through a single decoder-only Transformer. Why it matters: Bypassing latent compression cuts down pipeline complexity and proves a unified raw-pixel approach is entirely feasible.

  7. Reach 48

    CopilotKit open-sources OpenDots, an agent workspace with dedicated computers

    CopilotKit just launched OpenDots, an open-source template for running persistent AI agents on your own infrastructure. Each agent gets its own sandboxed browser, terminal, and workspace, switching between text chat, realtime voice calls, and Slack. Why it matters: it offers developers a self-hosted alternative to proprietary assistants, keeping full control over tools and sensitive data.

  8. Reach 48

    Raven: an AI agent manager that builds games and upgrades itself

    EverMind-AI open-sourced Raven, a self-improving meta-harness designed to orchestrate both built-in and third-party agents like Claude Code. Over a four-day autonomous run, it built a playable Godot 4 FPS game alongside its landing page and promotional posters. Why it matters: by letting AI refine its own planning and tool execution against locked benchmarks, it brings recursive self-improvement into practical engineering.

  9. Reach 48

    World Play 2 drops a real-time world generator with long-term memory

    World Play 2 just open-sourced an interactive world model that lets you roam and interact with characters and objects in real time. Instead of degrading over long horizons, it compresses past events into memory tokens to keep environments consistent even when you leave and return. It offers a 4-step Turbo version, though the full unquantized model weighs 69GB.

  10. Reach 59

    InSpatio-World 1.5 turns any image into an explorable 4D world

    InSpatio just open-sourced InSpatio-World 1.5, a spatial model that generates roaming videos from a single image, panorama, or video based on custom camera paths. It maintains spatial consistency and motion across time, and at under 6GB based on Wan 1.3B, it fits easily on consumer GPUs.

  11. Reach 48

    PDMD open-sources 4-step video distillation for MiniMax-H3

    Researchers just open-sourced PDMD, a video distillation technique that shrinks MiniMax-H3 generation from 30 steps down to just four. By curbing misleading critic feedback during training, it keeps motion smoother and visual details sharp instead of muddying the frame. Fast open-source video generation is getting shockingly usable.

  12. Reach 48

    Fermion drops Phonon-2, a 164MB speech model that beats giants

    Fermion Research just open-sourced Phonon-2, a lightweight English speech recognition model under 164MB. It transcribes an hour of audio in roughly 20 seconds on an M5 MacBook Air while scoring a 5.21% average word error rate. It proves high-accuracy ASR no longer needs multi-gigabyte models or cloud dependencies.

  13. Reach 59

    Nvidia puts AI agents inside a hardware cage

    Nvidia just launched the Open Agent Safety Platform, pairing open-source OpenShell runtime software with Sentry, a BlueField-4 DPU watchdog that quarantines misbehaving agents in milliseconds. The system moves guardrails out of the model and enforces zero-trust boundaries at the silicon level. Why it matters: as agents get autonomous system access, runtime security needs an unbypassable kill switch.

  14. Reach 48

    Run 100B+ models on a 12GB GPU with Strata

    Strata just open-sourced a cross-platform inference engine that runs massive models like Qwen3.8-Flash-Next on consumer PCs with just 12GB of VRAM. It tiers execution across your GPU, RAM, and SSD while exposing local OpenAI and Anthropic-compatible APIs. Why it matters: massive frontier models are officially leaving the cloud and moving straight onto everyday desktop hardware.

  15. Reach 48

    Magpie lets you run Claude Code on DeepSeek from the menu bar

    Open-source tool Magpie just turned the messy world of coding agent configs into a clean menu bar app. It sets up a local gateway that lets you plug models like DeepSeek and Kimi into Claude Code or Codex, automatically fails over when you hit rate limits, and even scrubs API keys and passwords before prompts leave your machine.

  16. Reach 48

    Valen brings fast-twitch visual reflexes to AI agents

    Valen-Team just open-sourced Valen, a lightweight multimodal model that brings System One fast-twitch decision-making to images and video. Instead of generating slow text explanations, the Qwen3.5-powered models (0.8B to 4B) directly output probability scores across predefined action candidates. It gives GUI agents, game bots, and robots a split-second reflex layer without massive compute overhead.

  17. Reach 48

    CLM-8B brings ultra-fast System 1 thinking to AI agents

    Contrastive-LM just open-sourced CLM-8B, a lightweight System 1 decision model trained to connect states and actions via contrastive learning. By caching state and action embeddings independently, it matches Jev on computer-use and tool-calling tasks with up to 9× lower latency, while scoring 87.6% on Terminal-Bench 2.1 as an agentic verifier.

  18. Reach 48

    Ming Image 0.1 turns flat designs into editable layers

    inclusionAI just open-sourced Ming-Image-0.1, a 6B model series purpose-built for visual design like UIs and posters. It natively outputs up to 2K graphics with transparent alpha channels, and a companion model can break flattened designs back into editable layers. Why it matters: It directly connects generative AI with real graphic design pipelines.

  19. Reach 48

    SupraLabs debuts Supra2-IMG, a 100M-parameter image generator

    SupraLabs just open-sourced Supra2-IMG, a tiny diffusion transformer with only 104M parameters and a footprint around 417MB. It was trained on 5.6M synthetic images in just 9 hours on a single H100 GPU. While resolution is capped at 256x256, it proves practical text-to-image models can comfortably squeeze onto phones and edge devices.