skainet

SKaiNET aims to democratize "Edge AI / On-device AI" by bridging the gap between high-level application development and low-level hardware optimization. We believe AI should be portable, type-safe, and developer-friendly, enabling seamless intelligence in everything from mobile apps to IoT devices without sacrificing performance.

For architecture details see ARCHITECTURE.md.

Add the core dependencies (Gradle Kotlin DSL):

dependencies {
    implementation("sk.ainet.core:SKaiNET-lang-core:0.16.0")
    implementation("sk.ainet.core:SKaiNET-backend-cpu:0.16.0")
}

Java / Maven users — see Java Getting Started for BOM setup and JVM flags.

val model = nn {
    input(28 * 28)
    dense(out = 128)
    relu()
    dense(out = 10)
}

val ctx = DirectCpuExecutionContext()
val ingestion = LlamaIngestion(ctx)
val weights = ingestion.load { SystemFileSystem.source(Path("model.gguf")).buffered() }
val tokenizer = GGUFTokenizer.fromSource(SystemFileSystem.source(Path("model.gguf")).buffered())

val runtime = LlamaRuntime(ctx, weights)
runtime.generate(tokenizer.encode("Once upon a time"), steps = 64) { token ->
    print(tokenizer.decode(token))
}

More examples: SKaiNET-examples | SKaiNET-notebook

• Targets: JVM, macOS (Native), JS, WASM (Browser + WasmWasi)
• Single codebase shared across all platforms via Kotlin Multiplatform

• KLlama: Llama-family models from GGUF files with streaming generation
• KGemma: Gemma 3n models from SafeTensors with HuggingFace tokenizer
• KBert: BERT and Sentence-Transformers for embeddings
• JVM acceleration via MemorySegment tensors, SIMD GEMM, paged KV cache

• Function / tool calling via skainet-kllama-agent
• AgentLoop for multi-turn tool-use conversations
• Java facade: JavaAgentLoop

• Sequential: nn { input(); dense(); relu(); dense() }
• DAG / Graph: arbitrary wiring with dag { } for ResNet, YOLO-style architectures
• Layers: Dense, Conv1d/2d/3d, MaxPool, AvgPool, BatchNorm, Dropout, LeakyReLU, ELU
• KAN (Kolmogorov–Arnold Networks) layer (experimental)
• Autograd engine with reverse-mode gradients, SGD and Adam/AdamW optimizers

• Built-in loaders: MNIST, Fashion-MNIST, CIFAR-10
• Formats: GGUF, ONNX, SafeTensors, JSON, Image (JPEG, PNG)
• Type-safe transform DSL: resize, crop, normalize, toTensor

• SKaiNET entry point, TensorJavaOps, builder-pattern model definition
• KLlamaJava / KBertJava facades for blocking and async inference
• Maven BOM (sk.ainet:skainet-bom) for one-line version management
• Docs: Getting Started | LLM Inference | Model Training

• Export trained models to standalone, optimized C99 with static memory allocation
• Ready-to-use Arduino library output
• See arduino-c-codegen.md

• Lower Kotlin DSL to MLIR StableHLO dialect
• Optimization passes: constant folding, operation fusion, dead code elimination
• Valid IREE-compilable output with streaming API and public HloGenerator
• See hlo-getting-started.md

• Unified LLM core — deduplicated and optimized implementation of KvCache, softmax, RoPE, and sampling
• Infrastructure cleanup — extracted core LLM and transformer code to standalone repositories SKaiNET-transformers) to streamline the core project
• Consistent BOM — refactored skainet-bom to use local project() references for reliable builds

See CHANGELOG.md for the full release history.

• Q1 2026: Comprehensive documentation ✅
• Q2 2026: Reference-based validation of computation correctness
• Q3 2026: Agentic AI enhancements ✅ (tool calling shipped in 0.13.0; ongoing)
• Q4 2026: Federated learning support for multi-device training

We love contributions! Whether it's a new operator, documentation, or a bug fix:

1. Read our Contribution Guide.
2. Check the Good First Issues.
3. Open a discussion or issue on GitHub.

Browse the full codebase documentation on DeepWiki.

• Dhia Chemingui (@dhiaspaner) — Android KMP plugin migration (#385, #386)

MIT — see LICENCE.