What’s new.

You can now generate images and video entirely on your machine, with no API key. We added a new local provider, GGML, built on stable-diffusion.cpp. Download open models like Z Image Turbo, Qwen Image, Wan 2.2, and LTX and start generating.

Resources:Z Image TurboFlux 2 Klein 4bQwen ImageWan 2.2 TI2V-5BWan 2.1 T2V 1.3BWan 2.2 T2V A14BLTX-2.3 DistilledH3 Local

You can now turn text into speech entirely on your machine. We added Supertonic to the local models through the HuggingFace runtime, with ten voices and a sample of each to play before you choose.

Resources:Supertonic TTS

If you run Ollama on another computer, you can now point CSuite at it. Open Advanced configuration on the Ollama entry in Models and enter the server URL. The same option is in onboarding, and a reset takes you back to the bundled server.

We added Meta's Muse Image to the image models on Runware. It keeps text in an image clean, blends up to 10 reference images into one composition, and costs about a cent an image.

Resources:Muse Image

Image›Muse ImageMuse ImageRunwareAbout a cent per imageProduct still life of a brushed-titanium espresso tamper standing on wet black slate, a scatter of coffee beans around it.GenerateReal output · the first-run sample from the model page

We added Google's original Nano Banana pair on Runware and Replicate. Pro renders at up to 4K, and its 2K size costs the same as 1K.

Resources:Nano BananaNano Banana Pro

You can now continue a clip instead of regenerating it. Extend works from the whole video or from just its last frame, and gives you back only the new footage.

You can now drop an empty image, video, or audio layer straight onto the timeline and generate into it with AI.

We added Google's Gemini Omni Flash 1.1 on Runware and Replicate. It makes 3 to 10 second clips, takes a first and a last frame, and can extend a clip you already have.

Resources:Gemini Omni Flash 1.1Gemini Omni Flash

You can now render Seedance 2.5 at 1080p on Runware, for about eight percent less per second than before. The 4:3, 1:1, 3:4, and 21:9 ratios work again too.

Resources:Seedance 2.5

We added Alibaba's Wan 3.0 Prime on Runware, the quicker tier of Wan 3.0. It renders a clip in roughly half the time for about 1.4 times the price. Everything else matches the base model: 2 to 30 seconds at up to 1080p, first and last frame control, reference media, and native audio.

Resources:Wan 3.0 PrimeWan 3.0

We added IBM Granite 4.2 to the local models through Ollama in 3B, 8B, and 30B sizes. All three call tools and read a 128K context. The 3B is a roughly 2.2 GB download that runs on 8 GB of memory, so there is a size for a modest laptop as well as a workstation.

Resources:Granite 4.2 3BGranite 4.2 8BGranite 4.2 30B

Model lists now open with the most recent releases, in the model pickers, the provider lists under Models, and the workflow model node. When two models suit your request equally well, Auto now picks the newer one.

We added Ornith 1.5 to the local models through Ollama in 9B and 35B sizes. Both read images and call tools, so you can use them in chat and workflows as well as the text workspace.

Resources:Ornith 1.5 9BOrnith 1.5 35B

When your model is set to Auto, you can now type settings straight into the prompt. "Red car, 16:9" picks the aspect ratio, and "8 seconds, with audio" sets a clip's length and sound. We only apply settings the chosen model supports, and only for that run, so your defaults stay as they are.

The wand that rewrites a rough prompt now looks at your selected model and anything you have attached. It keeps the subject of your song, writes real lyrics for models that sing from the prompt, writes an editing instruction when you attach an image, and describes motion when you set a start frame.

You no longer have to choose a model yourself. Every model picker now has an Auto option that reads what you are asking for and picks a model to match, so a logo goes to a model that returns a vector file and a quick draft goes to a fast one. New installs start on Auto everywhere, and if the best model is on a provider you have not connected, we offer the best one you can run.

We added Alibaba's Wan 3.0 on Runware. It makes clips from 2 to 30 seconds at up to 1080p, with first and last frame control, reference images, and native audio.

Resources:Wan 3.0

Onboarding and the License panel now show your real license status and the reason behind it. If your license has expired you can enter a new key, and a disabled Generate button now tells you why in a tooltip.

We went through the whole app and fixed around ninety issues, from editing and file handling edge cases to cloud generations being billed twice on a retry.

We added four Recraft image models, two of which are vector models that give you real SVG files. Z.ai's GLM-5.2 joins the text models, and Veo 3.1 Lite gives you a cheaper Veo tier on both providers.

Resources:Recraft v4.1Recraft v4.1 ProRecraft v4 VectorRecraft v4 Vector ProGLM-5.2Veo 3.1 Lite

More models now work with the key you already have. Qwen Image 3, Qwen Image 3 Pro, and Grok Imagine Image 2.0 are now on Replicate, and Kimi K2.6 and Grok Text-to-Speech are now on Runware.

Resources:Qwen Image 3Qwen Image 3 ProGrok Imagine Image 2.0Kimi K2.6Grok Text-to-Speech

We fixed images disappearing when you asked a Gemini chat model on Runware for one. The image now stays in the thread, and if something fails later in a turn you keep whatever it already made.

A .deb install cannot update itself, so CSuite now tells you when a new version is out and links you straight to the download page.

You can now remove the background from a video. The video AI menu writes a transparent WebM, so your clip can sit over other layers in a composition. We also moved the AI actions to the top of that menu.

We now ship a .deb package alongside the AppImage, so CSuite installs on Debian and Ubuntu with a proper desktop entry and its own icon in the dock.

Once an update has downloaded, you now see a Restart to update pill in the title bar on macOS. CSuite also checks for updates when you come back to the window.

We fixed local image generation. Ollama removed it in a recent release, so CSuite now bundles a second runtime pinned to the last version that supports it, downloaded only when you install an image model.

We fixed the Stop button on local generations. It used to cancel cloud work only, and now it cancels both.

You can now describe the workflow you want and the assistant builds or edits the graph for you, the same way the composition chat edits a timeline. Each edit is a single undo step.

We added three models: MiniMax H3 on Replicate, Gemini 3.6 Flash on Runware, and Qwen3.8 27B to run locally through Ollama.

Resources:H3Gemini 3.6 FlashQwen3.8 27B

You can now attach reference images, videos, and audio wherever a model supports them, on both cloud providers. A picker in each composer shows you exactly what the selected model accepts.

You can now attach PDF, text, Markdown, and HTML files to a chat, and their text is read into the conversation. Cloud models that support vision now accept your images too.

You can now ask Chat to run one of your saved workflows by name, and it posts the results back into the thread as attachments.

Your title clips can now use any font installed on your machine, searchable in the picker. The composition assistant also knows which fonts you have when it writes a title.

Stop in Chat now cancels the work in progress instead of only hiding it, and every round has a timeout. You can copy a single message or the whole conversation as Markdown, regenerate a reply, and edit and resend your last message.

Workflow runs now cancel for real, and a failed node keeps its error and offers a Retry. The toolbar estimates the cost before you run, and we added auto-layout, checks before a run, nine new edit operations, and the option to save your own graph as a template.

The dashboard now opens with six create tiles and a Files panel grouped by type, including your compositions, workflows, and chat threads. Click any card to open it in the workspace it belongs to.

You can now send a still on the composition timeline to a video model and have the result replace it, without leaving the editor.

We added Lightricks LTX-2.5 to the video models, in a fast tier and a pro tier.

Resources:LTX-2.5 FastLTX-2.5 Pro

The rebuilt composition editor now does everything the old video editor did, so we removed the old one. Compositions made in the old editor are not carried over, so export anything you still need from them first.

You can now switch a composition between square, landscape, and portrait, and every frame reflows to the new shape.

We added Grok Imagine Image 2.0 to the image models, and Muse Glimmer and Nemotron 3.5 Lightning to run locally through Ollama. Runware image models also moved to their 2K tier.

Resources:Grok Imagine Image 2.0Muse GlimmerNemotron 3.5 Lightning

We rebuilt the video composition editor. You get tracks you can lock, mute, and hide, drag and drop onto the canvas, and clips you can move and resize directly. An AI chat edits the timeline through real undoable steps and can generate assets for you, and you can export MP4 at the resolution and framerate you choose.

We added ByteDance's Seedance 2.5 on both Replicate and Runware.

Resources:Seedance 2.5

We fixed audio trim looking like it did nothing. The result was being saved in a different format from the file you were looking at.

You now get a system notification when a generation finishes while CSuite or that workspace is not in front, and clicking it takes you straight there. You can turn this off in Account.

We fixed Stop on Runware generations. The button used to keep spinning instead of ending the run.

We fixed background removal on Replicate, which was failing because the underlying model had moved. Community tool models are now pinned to a known version.

We fixed several Runware voice models and the music models, which were failing because a default the provider expects was not being sent.

The wand that turns a rough prompt into a fuller one is now in the text and video workspaces as well as image and audio.