Open Source

Hasantoxr documents AgenticSeek with Ollama, SearXNG, and Docker, including install steps, model config, and a locked WORK_DIR. The stack keeps models, chats, and files on the user's machine.

Wan-Streamer v0.1 surfaced with paper links and demos showing real-time video conversations, live recording, and spoken avatar responses. That matters for interactive characters and live creator tools because multimodal generation moves from rendered clips to low-latency back-and-forth.

Builders published GGUF conversions, loader nodes, and local benchmarks for Krea 2 Turbo after yesterday’s open-weights release, alongside new multi-style and watercolor tests. The follow-up matters because creators now have clearer ways to run, tune, and style-push Krea locally on smaller VRAM budgets.

Krea 2 Turbo arrived with open weights, commercial rights, ComfyUI workflows, and community GGUF and FP8 ports that users say can run locally on modest hardware. Early benchmarks praise speed and style range, while some testers flag drift toward recognizable IP.

Ideogram 4.0 shipped as an open-weight image model with JSON prompting, bounding boxes, stronger text rendering, and native 2048px output. The release targets layout-heavy creative work, and teams can test early fal and Leonardo integrations in production flows.

Runway joined NVIDIA's new Cosmos Coalition as a founding member, and Runway says the group's first project is a base model it is codeveloping with NVIDIA. NVIDIA also says Cosmos 3 is fully open with weights and post-training recipes, so teams can track the shared world-model stack.

Presenton was highlighted as an open-source presentation generator that turns prompts and documents into editable PPTX or PDF decks and can run self-hosted with BYOK and Ollama. Use it if you want slide generation outside locked SaaS editors while keeping exportable files.

Posts report SenseTime open-sourced SenseNova U1, a unified text-image model with interleaved generation, 8-step distilled LoRA and ComfyUI workflows. They cite 2K image times around 15 seconds and H100 inference cuts to about 2 seconds, so compare it against your current image pipeline.

A community post spotlights OmniVoice Studio, an open-source local dubbing pipeline that transcribes, translates, clones voice from 3 seconds and remixes dubbed audio back into video. Running locally keeps voice data on device and removes subscription costs, so it may fit privacy-sensitive dubbing workflows.

BenchLocal v0.2.6 adds reachability checks so offline local models are skipped and resumed instead of breaking side-by-side tests. The update is aimed at tight-VRAM setups where creators and tinkerers load providers one after another on the same machine.

Supertone open-sourced Supertonic, a local TTS engine that runs faster than real time on phone CPUs with ONNX models and cross-language runtimes. Voice apps and audiobook workflows can use it to avoid per-character API billing and keep audio generation private.

Harbor 0.4.18 added one-command access to Open Design and Voicebox, bundling a local-first design app and a voice cloning and TTS studio inside one homelab layer. The release cuts setup friction, so users can migrate both tools into a single local install path.

Posts show an open-source toolkit that turns one reference image into an interactive 3D scene with generated meshes, lighting, physics, and sound. The demo stack chains World Labs, Hunyuan 3D, ElevenLabs, and fal rather than a single native model.

Recordly launched as a free open-source alternative to Screen Studio with auto-zoom, cursor effects, local editing, MP4 and GIF export, and an extension marketplace. It matters for tutorial and product-demo creators because capture and export stay on-device, though current evidence is still mostly repo promotion and reposts.

An 80.lv-cited report said Anthropic paid €240K to sponsor Blender, while Blender leadership said the deal was not an AI takeover and creator posts said the public patronage was later pulled. The connector still appears to ship, but the funding tie-in became a flashpoint inside open 3D communities.

Posts describe MeiGen as a free prompt gallery for GPT Image 2, Nano Banana 2, Seedance 2.0, Veo 3.1 and Midjourney, with drag-to-canvas generation and reverse prompting. The thread says the dataset is open source and already wired into Claude, Figma and OpenClaw workflows.

Posts say LTX HDR can generate HDR video or convert SDR footage to scene-linear EXR through API, ComfyUI and Hugging Face. The beta matters because EXR output is described as workable in Resolve and Nuke, but the release is still tagged v0.9.

DeepSeek V4 Preview surfaced as an open-source 1M-context model family, with early docs and community testing pointing to Flash and Pro variants. The release matters for creators and vibe coders looking at self-hosted options, but most performance claims are still coming from first-wave community benchmarks.

Google published the draft DESIGN.md specification so colors, typography, components, and rules can live in one AI-readable file, with a CLI validator and components support in progress. That matters because design agents and handoff tools can point to one structured source of truth instead of inferring UI rules from scattered docs.

OpenGame released an open-source agent that turns prompts into playable web games, with public demos spanning shooters, quiz battlers and 90s-style fighters. The release matters because game ideas now arrive as runnable browser prototypes rather than static mockups, though the current proof points are demo-heavy.

Stable Diffusion and VFX creators say LTX 2.3's distilled LoRA v1.1 improves motion and custom-audio sync. Posts show local short-film and flight-shot workflows running through ComfyUI and Resolve on consumer GPUs.

Tencent released HY-World 2.0 with WorldMirror 2.0 code and weights for turning text, images, or video into persistent 3D scenes. The output includes navigable geometry and camera data instead of disposable video frames.

OpenBMB released VoxCPM on GitHub with text-described voice design, 3-second cloning, 48kHz audio, and 30-language support. The Apache 2.0 release makes multilingual voice work and local self-hosting cheaper.