Models
Stacknet provides access to models through Model Layers, fine tuned pools of models and capabilities. Models are served by nodes operated by aISPs.
Available Model Layers
Layers (recommended)
Layers are the canonical names you should pass as model in /v1/chat/completions. The network picks a member model that best fits your prompt + content type.
| Model | Modalities | Capabilities | Context |
|---|---|---|---|
preview | Text, Image → Text | Function calling, Structured output, Reasoning | 128K+ |
duce | Text, Image → Text, Image → Image, Text → Video, Image → Video, Video → Video, Text → Music, Music → Music, TTS | SOTA music, Function calling, Reasoning, MoM | 128K+ |
magma | Text, Image → Text, Image → Image, Text → Video, Image → Video, Video → Video, Text → Music, Music → Music, TTS | SOTA video, Function calling, Reasoning, MoM | 1m |
Capability aliases
/v1/models also returns four capability-scoped aliases. They’re lower-level than the layers above — each maps directly to a specific underlying model rather than a routing tier. Use them when you need a fixed capability shape without layer-routing semantics.
| Alias | Capability | Notes |
|---|---|---|
stack-chat | Text → Text chat | Conversational completion. |
stack-code | Code generation / repair | Optimised for code edits and tool-use loops. |
stack-poly | Multimodal in/out | Vision + audio understanding; text/image/audio output. |
stack-embed | Embeddings | Use against /v1/embeddings, not /v1/chat/completions. |
If you’re starting fresh, prefer the layer names — the layer router handles routing decisions for you and inherits any future capability gains. The stack-* aliases are kept for backwards-compatibility with earlier integrations.
Creating a Model Layer
Anyone on the network can create network wide supported model layers and earn every time the model layer is used. A model layer will qualify and propagate if the following requirements are met:
- 90% network support (can run on the hardware of the network. Making hardware targeting moot. Real network-wide performance is the bar.)
- 40% different composition from the leading model layer
- 2.5%+ better benchmark than the leading model layer
- (Cumulative model parameters x 10 ) + Network Benchmark tokens = Token Cost
- Written and compiled in Lode Language
Try a Model Layer
Visit the Chat API playground to test model layers interactively.