Home › Models
This is the page for the awkward questions. What actually runs well on device today? What will my laptop do? Can I sell what comes out of it? Short answers, in that order.
If you only read one section, read maturity by modality — it is the difference between a good afternoon and a frustrating one.
SeasonStudio does not invent its own inference engine. It leans on two well-travelled local stacks, and which one a panel uses determines what file format you feed it and which backends are available.
| Panel | Runtime | Model format |
|---|---|---|
| Text, Translate | llama.cpp / GGML | .gguf, usually 4-bit quantised |
| Music, Text to speech, Speech to text | GGML | .gguf |
| Image | stable-diffusion.cpp / ONNX Runtime | Model file plus a matching VAE |
| Video | ONNX Runtime | Model files plus a shared text encoder |
| Vision | ONNX Runtime | .onnx |
Each runtime enumerates what your machine offers and takes the fastest option it can use. CPU always works and is always available; the accelerated backends are part of the one-time Premium unlock.
| Platform | GGML backends | ONNX Runtime providers |
|---|---|---|
| Windows 10 / 11, x64 | CPU · Vulkan · CUDA | CPU · DirectML · CUDA |
| macOS, Apple silicon | CPU · Metal | CPU · CoreML |
CUDA is the fast path on NVIDIA hardware; DirectML covers AMD, Intel and NVIDIA alike on Windows; Metal covers Apple silicon. On a machine with none of them the app still runs — it simply asks for more patience.
Open, locally runnable models did not all arrive at the same time or at the same quality. Text and image generation have had years of iteration; on-device music and video are much younger. Here is where we think each panel genuinely stands, and why.
Quantised instruct models in the 1‑to‑8‑billion parameter range are the most settled part of local AI. They draft, summarise and restructure reliably at paragraph and page scale. Dedicated translation models go further still: for major language pairs they hold register and terminology better than a general model of similar size. The honest limits are knowledge and reasoning depth — a small local model does not know recent events, and will produce a confident wrong answer rather than admit uncertainty.
This is the modality where local and hosted quality have converged most. The SDXL and SD 3.5 lines, and the newer families alongside them, produce illustration and concept work you can put in front of a client. Prompt adherence has improved substantially, editing models let you change one thing without redrawing everything, and the sampler preview means a bad seed costs seconds rather than minutes. What remains hard: legible text inside the image, hands and crowds, and exact compositional control.
Transcription is a solved problem for clean audio, and it is the task where running locally is most obviously worth it — recordings are the most sensitive thing most people feed a model. Synthesis has caught up quickly: current small speech models are natural enough for narration, and describing a voice in words rather than picking from a list genuinely works. The seams show over long passages, where prosody loses the thread, so work paragraph by paragraph.
The vision stack is made of small, mature, purpose-built models rather than one ambitious one, which is exactly why it is dependable. Captioning, OCR, classification, detection and segmentation all behave predictably on ordinary photographs and load in a second. Accuracy falls off with handwriting, stylised type and heavy compression, and detection can only name classes it was trained on.
The Stable Audio family made short-form audio generation practical on a desktop, and within its range the results are genuinely usable: effects, textures, ambient beds and loops. Beyond that range the field simply has not arrived. Long-form musical structure — a theme that develops, a chorus that returns — is not something the current generation of open models does, and vocals are out of scope. Our advice is to use it for the layer underneath rather than the tune on top.
Local video generation is the most exciting and least dependable thing in the app. The open text-to-video and image-to-video models can produce a few seconds of coherent motion, and occasionally something lovely, but the cost is high: gigabytes of model files, minutes per render, modest resolution, and identity that drifts across frames. We ship it because it is real and improving fast, and we label it experimental because we would rather set your expectations than sell you a surprise.
Because the alternative is you discovering it on a deadline. The gap between modalities is the single most useful thing to know about on-device creative AI, and it is the thing marketing pages usually leave out.
The tiers below are our own expectations from development machines, not results from a broad measurement pass. Treat them as a guide to which column you are in, not as a benchmark. They will be replaced with measured numbers.
Find your machine in one of the three tiers, then read across. The point of the table is not exact timings — it is knowing which panels to open first.
| Tier | A machine like this | What to expect |
|---|---|---|
| Light | Thin laptop, 8 GB RAM, integrated graphics or a base Apple silicon machine. | Text, Translate, Speech to text, Text to speech and Vision are all comfortable. Image works at 512×512 with a light model — expect tens of seconds per picture. Music is slow but usable for short clips. Skip Video. |
| Comfortable | 16 GB RAM with an 8 GB graphics card, or an Apple silicon Pro machine with 16 GB or more unified memory. | The sweet spot. Image at 1024×1024 in well under a minute, larger language and translation models, longer transcriptions, Music without waiting around. Video becomes possible for very short clips at low resolution. |
| Roomy | 32 GB RAM with a 16 GB or larger graphics card, or an Apple silicon Max machine. | Everything, including Video at the resolutions the models were actually designed for. Heavy image models, batches of variations and the largest language models all fit. |
One compact multimodal language model is built into the app, so the Text panel and the content safety gates work out of the box with nothing downloaded. Every other model is yours to add — either fetched on demand when a run needs it, or imported from a folder you already have.
| Panel | Families it can load | Rough size on disk |
|---|---|---|
| Text | Compact instruct models, and larger ones in the 4‑to‑8‑billion class | 0.5 – 6 GB |
| Translate | Dedicated translation models, small and large | 1 – 5 GB |
| Image | SD 1.5, SDXL, SD 3.5, newer open families, and editing models — each with a VAE | 2 – 20 GB each |
| Video | Text-to-video and image-to-video models plus a shared text encoder | 10 – 40 GB |
| Music | Stable Audio, in effects and music variants and two sizes | 1 – 5 GB |
| Text to speech | Speech models in base, custom-voice and voice-description variants | 0.5 – 2 GB |
| Speech to text | Whisper-family models across several sizes | 0.1 – 3 GB |
| Vision | OCR, classification, detection, segmentation and face models | 10 – 300 MB each |
The exact list of models the app offers moves as the field does. The guide explains how to add one you already have.
SeasonStudio is a player, not a publisher. The app's own licence covers the app; every model you load stays under the licence its authors chose, and that licence — not ours — decides what you may do with what comes out.
Open model licences vary more than people expect: some are permissive, some are community licences with conditions, and a few are research-only. The app lists the licence next to every model it can load. Read it once, before the client work, not after.
The inference stacks themselves — llama.cpp, stable-diffusion.cpp, GGML and ONNX Runtime — are open source under permissive licences. On Windows, the NVIDIA CUDA and cuDNN runtime libraries are redistributed under NVIDIA's own terms and are not covered by ours. The full third-party notice list ships with the app.
We claim nothing in what you make. Nothing is uploaded, so nothing is collected, and there is no clause anywhere granting us a licence to your output. What you generate is a file on your disk.