Runware

Runware’s MCP Server is live 🤖 The full Runware API is now available as an MCP server. Claude, Cursor, Codex, and any other MCP-compatible AI tool can call Runware directly, just by chatting with your agent. • Image generation, with advanced parameters • Background removal, upscaling, captioning, masking • Audio, video, 3D, and LLM inference • Model search across the full Runware catalogue Runware in the agentic world starts today. Launch link in the comments.

Seedance 2.0 Mini is now on Runware 🎬 It's the lightweight tier in the Seedance 2.0 family, based on the same architecture that sits at the top of the Artificial Analysis text-to-video and image-to-video leaderboards, trimmed down for speed and volume. It takes text, reference images, or a reference clip and returns a coherent cinematic short, with consistent characters and basic camera control, at the lowest price point in the Seedance family. Up to 15 seconds at 720p, native audio-video generation, multimodal reference input. Launch link in the comments.

Fish Audio S2.1 Pro is now on Runware 🎙️ Fish Audio's flagship TTS model brings highly expressive, low-latency speech generation to your pipelines. You get 80+ language support with automatic language detection, inline bracket cues for emotion and delivery control (think [laughs], [whispering], [short pause]), and multi-speaker dialogue in a single generation call, with no manual stitching required. The model streams audio with very fast time-to-first-audio and supports zero-shot voice cloning from a reference clip. Up to 4 output variations per call, with adjustable speed, temperature, and loudness normalisation. Launch link in the comments

We teamed up with NightCafe for an AI art competition for all skill levels. Submit a project that tells a complete story from start to finish. Expert judges will review every entry, and the winner takes home a $1,000 grant plus a feature in Prompt Magazine. Ready to tell your story? https://shorturl.at/h2gj6

HappyHorse 1.1 is now on Runware on D0 🐎 Building on 1.0, version 1.1 reworks the parts that caused friction in production: stiff motion is replaced with fluid, weighted actions; oily, over-sharpened faces give way to natural skin textures that hold up in close-up shots; and audio moves beyond basic sync to emotional dialogue delivery with precise A/V timing. Instruction following also gets a significant upgrade. A single prompt can now describe 6 - 8 consecutive scenes with automatic camera dispatch, with stronger consistency across multi-character scenes and storyboard layouts. Launch link in the comments

HappyHorse 1.1 is now on Runware on D0 🐎 Building on 1.0, version 1.1 reworks the parts that caused friction in production: stiff motion is replaced with fluid, weighted actions; oily, over-sharpened faces give way to natural skin textures that hold up in close-up shots; and audio moves beyond basic sync to emotional dialogue delivery with precise A/V timing. Instruction following also gets a significant upgrade. A single prompt can now describe 6 - 8 consecutive scenes with automatic camera dispatch, with stronger consistency across multi-character scenes and storyboard layouts. Launch link in the comments

Last Friday, a US government export-control directive hit Anthropic. By that evening, Fable 5 and Mythos 5 were switched off for every customer worldwide, days after Fable 5 went public. Not an outage. A model which teams had already built on was simply gone. The lesson isn't open versus closed. It's the risk of betting your product on one model you don't control, when it can be switched off or repriced overnight. And the case for open keeps getting stronger. It now does most production work at a fraction of the price: MiniMax M2.7 runs at $1.20 per million output tokens, GPT-5.5 charges $30. For most of the job the quality gap no longer shows. Closed keeps a real edge on only the hardest tasks. The teams that came through Friday fine weren't relying on one model. They run their open models behind one API and route each task to whatever fits, so swapping one out is a config change, not a rebuild. The model is becoming the easy part. The advantage now lives in the stack underneath it. Friday was the symptom. We wrote about the bigger shift behind it: https://lnkd.in/e-QBb5RW

P-image-try-on is now on Runware 👕 Getting product-on-person imagery usually means a photoshoot. Skip that entirely with P-Image-Try-On from Pass a person photo and up to eleven garment reference images. The model realistically dresses the person in all of the provided items. No text prompt required; the garment images are the input. The use cases are broad: fashion teams generating product imagery, styling apps serving outfit previews on demand, try-before-you-buy flows that need multi-garment combinations. Launch link in the comments

DeepSeek AI V4 Pro & Flash now on Runware 🧠 DeepSeek V4 Pro joins DeepSeek’s fast and efficient V4 Flash model on Runware, offering closed-source performance at open-source cost. DeepSeek's flagship V4 model sits at the top of the open-weight leaderboard across math, STEM, and coding benchmarks while matching the performance of the best closed-source models available today. For teams building coding assistants, long-horizon agents, or large-document pipelines, it's now accessible via the Runware API. Under the hood: a 1.6T parameter MoE with 49B active params, a 1M token context window, and up to 384K output tokens. That context scale makes it practical for large repositories, retrieval-heavy pipelines, and agentic systems. Launch links in the comments.

Ray 3.2 is now on Runware 🎬 The model covers text-to-video and image-to-video from a single endpoint. Pass a start frame, an end frame, or both to anchor the generation, and the model will hold visual continuity across the clip. The standout capability for production pipelines is native HDR output with optional EXR export in the ACES2065-1 standard. For teams working with broadcast or film color-grading workflows, that removes a post-processing step that most current video models don’t cover. Resolutions up to 1080p, across six aspect ratios including 21:9 cinematic ultrawide. Also supported: seamless looping for 5s clips, and forward or backward video extension from any prior generation. Launch link in the comments