Nace AI just released a 51-page technical report on Drex, its open-weight model family designed to pick the best answer from fixed options. Trained on 1.15 million verified decisions, it delivers calibrated probabilities in just 183ms. Open-weight versions are available on Hugging Face alongside a web test console.
Cactus just dropped Mirage, a tiny 51.3M-parameter vision model compressed to roughly 22MB via 2-bit quantization. It handles scene description, object detection, OCR, segmentation masks, and spatial reasoning in a single pass—clocking 74ms on a Galaxy S26 NPU. It proves full-featured visual perception can comfortably live on device edge chips without touching the cloud.
Video Rebirth just released HyperFlow H3 Flash, an open ~14B distillation of MiniMax-H3 that generates synchronized 768p video and audio in 4 to 8 steps. The Q4 GGUF weight drops down to 9.1 GB to run on 24GB RTX GPUs, while 4x Blackwell B200s achieve playback-speed real-time generation. Why it matters: Cutting-edge, synchronized audiovisual models are finally breaking out of hyperscaler clusters and onto local creator hardware.
Google just open-sourced Agile Deliberation, an AI framework that flips traditional prompt writing. Instead of forcing you to define subjective concepts upfront, the model actively hunts down tricky edge cases to help you pinpoint your own criteria—making content moderation and brand style labeling much cleaner.
Open-source project hunch turns Google's EmbeddingGemma 2 into a local, reflex-speed decision engine running directly on commodity CPUs. It acts as a fast student that learns from frontier LLMs and takes over repetitive chores like tool trace verification in just 22ms. Basically, it lets you slash agent inference bills without needing a GPU.
Kimi just released the open weights for K3.1 Pro on Hugging Face under a dedicated license. Billed as its most capable 3T-class model yet, the release packs native multimodality and agentic capabilities directly into the hands of open-source builders.
A two-person lab called Linum just open-sourced Pyramid-JiT, a 2B-parameter text-to-image model trained entirely from scratch for only $21,000. It ditches the standard VAE to generate images directly in pixel space, delivering solid visuals on a shoestring budget. While it struggles with text and complex anatomy, the Apache 2.0 release proves training diffusion models doesn't require millions.
Ricursive Intelligence just released Ripple-65B Preview, a dense open-weight model with a 1M context window trained for $32M. Skipping MoE complexity, it scored 80.52 on TerminalBench 2.1, beating massive frontier open-weight models over ten times its size in agentic coding.
Anthropic just open-sourced FlashProve, a provable inference system using cryptographic zero-knowledge proofs to verify an AI response actually came from the claimed model. Verifiers do not need weights or GPUs, checking proofs on a standard CPU in seconds, though generating proofs still takes 72 to 94 times more GPU compute.
Researchers just open-sourced NitroFusion 2, slashing image diffusion down from 8 steps to just 1 or 2 steps. By introducing adversarial distillation directly into the multimodal text stream, it preserves fine details and legible typography even in a single pass. The weights are available on Hugging Face and run locally via ComfyUI.
Pruna AI just released P-H3-Image, an open-weights image model built by reworking the MiniMax-H3 video model. By distilling components and slashing diffusion steps from 50 to 8, it runs 5x faster and shrinks down to 20 GB while rendering up to 16 megapixels. ComfyUI already supports it natively.
Sandflow AI just open-sourced sand3d, a Python library built on Blender that lets LLMs construct parametric 3D objects and robotic joints in readable code. The entire API fits in a 3,600-token prompt—using up to 6x fewer tokens than raw Blender script—and launches alongside a massive 100,000-object open dataset.
NVIDIA just open-sourced Auto Ontology, an Apache 2.0 framework that builds a semantic reasoning layer over enterprise databases. It translates ambiguous questions into validated, deterministic SQL queries with verifiable data provenance, and plugs directly into agents via MCP.
Moonshot AI just released the full open weights for Kimi K3.1 Pro, a 2.8-trillion-parameter MoE model that activates 104B parameters per token. It features a 1M token context window and native multimodal reasoning. Its massive 1.56TB size targets enterprise clusters running vLLM or SGLang.
StudyOnTV just turned screen time management into a learning gate for Android TVs and tablets, forcing kids to answer questions before earning viewing minutes. It runs completely locally with no cloud accounts, and can plug into AI models like DeepSeek to auto-grade subjective answers.
DeepSeek just open-sourced dsh-superagent-prototype, an experimental background agent designed to handle GitHub tasks without cluttering human attention. It runs tasks in isolated sessions and only pings developers when a pull request is ready or human intervention is strictly required. Keep in mind that DeepSeek frames this as research code rather than an official product.
Aura just open-sourced Autopilot, a real-time game playtesting agent for Unreal Engine 5, Unity 6, and Godot 4. Developers simply prompt it with an objective, and the agent reads project code, plays through levels, and visibly improves its skills over repeated runs while generating bug reports. It also connects directly to Claude Code and Codex via MCP.
Edge0 just open-sourced Audio8-TTS-Realistic, a lightweight 601M-parameter text-to-speech model supporting 23 languages. It pulls off zero-shot voice cloning from just a few seconds of clean reference audio and weighs only 2.6GB under an Apache 2.0 license. Why it matters: you can now run fast, commercial-friendly voice cloning on consumer-grade GPUs without massive compute.
CMU developer Zheyuan Kong just launched Auto-Company to cure the endless meeting plague among multi-agent demos. Fourteen persona-driven agents brainstorm and debate, but a hardcoded convergence loop forces a hard GO or NO-GO decision, triggering Claude Code to write and deploy real code by round three. It runs locally via Bash and Markdown, but requires full system access.
Pika just open-sourced Pika H3 Flex, a post-trained model built on MiniMax-H3 that generates videos with synchronized stereo audio in only six sampling steps. It handles 5-to-15-second clips across 11 aspect ratios, accepts up to 12 multimodal references, and runs on a single H100 GPU. Why it matters: professional creators get rapid, high-control iteration without waiting minutes for a render.
Nvidia just released Nemotron 3.5 Transcribe, an open-weight speech model that combines transcription, diarization, translation, and audio Q&A into one checkpoint. It supports 40 languages, handles hour-long audio, and identifies up to eight speakers. Thanks to a 30B Mamba-Transformer MoE architecture, it activates just 4B parameters per token while leading major open ASR benchmarks.
Sperid Labs just open-sourced Idol, a new model that turns a single 2D image into a standard 3D Gaussian splat with up to a million points. It handles background removal, accurately guesses the hidden backside, and caps out at just 7.2 GB of VRAM. Why it matters: High-fidelity 3D reconstruction is finally fast and lean enough for local desktop rigs.
Nvidia just launched Open H3 Regenerate 2K, an open-source tool that redraws low-resolution video drafts at 2K and up to 4K in just three diffusion steps. Built primarily for MiniMax H3 but compatible with models like LTX-2, it gives creators full local control over high-res video refinement, though it may slightly alter fine facial details.
NVIDIA just open-sourced MotionBricks, a real-time motion generation model for digital characters and humanoid robots. By feeding it a handful of target poses, it instantly connects the dots to generate smooth transitions and object interactions without retraining. Pretrained models for human skeletons and the Unitree G1 robot are ready to run on GitHub.
Developer Nutlope just open-sourced Bygga, a web app that turns flat IKEA paper instructions into interactive, animated 3D guides. It covers 1,171 products completely offline, reads steps aloud, and listens for voice commands so you can say "next" without dropping your screwdriver. Unknown manuals can be scanned and drafted via Together AI.
AI 3D platform Hi3D just open-sourced its V1.0 model weights and code under a commercial-friendly license. Sized at just 7.74GB, it runs on consumer GPUs to turn single images into detailed meshes up to 1536³ resolution. It hands creators a solid local baseline, even though the flagship V3 remains closed-source.
Baidu just open-sourced Unlimited OCR 2, a 3B document-parsing model that can ingest thousands of pages in a single request. By revamping attention with SR-SWA, it keeps memory and latency stable over massive documents, automatically stitching together cross-page tables and split text. It easily fits on consumer GPUs, meaning you no longer have to chop books into pieces just to digitize them.
Designer AmirMushich just open-sourced a clever technique to create interactive, rotatable web product pages without 3D modeling. The method generates a spinning video from reference images using LTX-2.5 Pro, splits it into about 30 frames, and controls playback via JavaScript based on cursor movement. Why it matters: E-commerce brands can build sleek pseudo-3D product showcases with zero 3D modeling overhead or runtime AI costs.
Tencent Hunyuan just open-sourced Hy-SimulMT, a bidirectional Chinese-English simultaneous translation model with roughly 3-second streaming latency. Beyond delivering real-time speech and text, it clones the speaker's vocal tone into the target language. The repository includes full inference and training tools, designed to keep multilingual meetings moving without awkward pauses.
MotionSites just open-sourced a Claude skill that builds cinematic, gallery-level landing pages instead of generic AI layouts. It packs 40 curated design prompts, a ready-to-use video and 3D media shelf, and an automated Playwright screenshot loop to catch and fix visual bugs. Why it matters: It replaces vague prompts with strict design rules, delivering web pages that actually look hand-crafted.
Bria just released open weights for Fibo 2.0 Edit Turbo, an 8B-parameter image editing model under a non-commercial license. Distilled down to just 4 steps without classifier-free guidance, it uses structured prompts to swap objects or alter styles while keeping untouched areas pixel-stable.
Pagekin is a newly open-sourced AI novel-writing agent designed to prevent story drift and broken character lore. Instead of dumping entire books into context, it structures world-building, character files, and chapter blueprints to load context on demand, offering granular scene rewrites and revision diffs. The tool is free on GitHub, but users must bring their own model API keys.
SenseTime just open-sourced SenseNova 6.8 Code2Video, a model that builds animations by generating executable code rather than diffusion sampling. Because every element is defined in code, you can tweak text, timing, or trajectories without re-rolling the entire clip. The 27B model weighs 55GB and requires a high-end GPU to run.
