SKaiNET Transformers

SKaiNET Transformers is a Kotlin Multiplatform inference engine for large language models. It loads GGUF and SafeTensors models, builds compute graphs from DSL network definitions, applies optimization passes, and executes inference on CPU (with SIMD acceleration) or GPU backends.

Key Features

Unified pipeline — load any supported model with a single CLI, auto-detected from GGUF metadata
Tool calling — agent loop with tool execution for any model that supports chat templates
Compute graph optimization — transpose elimination, weight deduplication, RMSNorm/SwiGLU/QKV fusion
Kotlin Multiplatform — runs on JVM, macOS Native, Linux Native, JS, and WASM
Quantization support — Q4_K_M, Q8_0, and other GGUF quantization formats with SIMD dequantization

Supported Model Families

Every generative family below with a "Verified" entry has a model-gated golden-token parity test asserting greedy-decode equality against a reference implementation, on the same DSL path the CLIs ship. The tests skip quietly when the model file is absent, and the fixture headers record the exact oracle build and commands.

Family Models Verified against Tool Calling DSL Network

Family	Models	Verified against	Tool Calling	DSL Network
LLaMA	LLaMA 2/3, Mistral	mainline llama.cpp — full greedy text equality (Llama-3.2-1B-Instruct Q8_0)	Yes	`llamaNetwork()`
Qwen	Qwen2, Qwen2.5, Qwen3, Qwen3.5	mainline llama.cpp — full greedy text equality on both variants (Qwen2.5-0.5B-Instruct Q8_0 with attention biases; Qwen3-1.7B Q8_0 with QK-norm)	Yes	`qwenNetwork()`
Gemma	Gemma 3 (FunctionGemma), Gemma 4. Gemma 2 has no supported path.	llama.cpp — golden-token chat parity (Gemma-4 E2B GGUF)	Yes	`gemmaNetwork()`
Gemma 3n	Gemma 3n E2B/E4B	mainline llama.cpp — full greedy text equality (gemma-3n-E2B-it Q4_K_M; AltUp, Laurel, sparsity, PLE, shared KV)	No	`gemma3nNetwork()`
Apertus	Apertus 8B	mainline llama.cpp — full greedy text equality (Apertus-8B-Instruct Q4_K_S; QK-norm, xIELU, ungated FFN)	No	`apertusNetwork()`
BitNet	BitNet b1.58 2B4T	bitnet.cpp and the HF BF16 reference — three-way token equality on the packed I2_S path	No	`bitnetNetwork()`
BERT	MiniLM, BERT variants, LEAF/BGE/E5 embedders	sentence-transformers (MongoDB/mdbr-leaf embeddings)	No	`bertNetwork()`
T5	gtr-t5-base (GTR embedder, vec2text inversion)	Real-weights gtr-base round-trip test	No	Hand-coded
Voxtral	Voxtral TTS	Not yet — architecture code only, no runtime facade	No	`voxtralBackboneNetwork()`

LLaMA

LLaMA 2/3, Mistral

mainline llama.cpp — full greedy text equality (Llama-3.2-1B-Instruct Q8_0)

Yes

llamaNetwork()

Qwen

Qwen2, Qwen2.5, Qwen3, Qwen3.5

mainline llama.cpp — full greedy text equality on both variants (Qwen2.5-0.5B-Instruct Q8_0 with attention biases; Qwen3-1.7B Q8_0 with QK-norm)

Yes

qwenNetwork()

Gemma

Gemma 3 (FunctionGemma), Gemma 4. Gemma 2 has no supported path.

llama.cpp — golden-token chat parity (Gemma-4 E2B GGUF)

Yes

gemmaNetwork()

Gemma 3n

Gemma 3n E2B/E4B

mainline llama.cpp — full greedy text equality (gemma-3n-E2B-it Q4_K_M; AltUp, Laurel, sparsity, PLE, shared KV)

gemma3nNetwork()

Apertus

Apertus 8B

mainline llama.cpp — full greedy text equality (Apertus-8B-Instruct Q4_K_S; QK-norm, xIELU, ungated FFN)

apertusNetwork()

BitNet

BitNet b1.58 2B4T

bitnet.cpp and the HF BF16 reference — three-way token equality on the packed I2_S path

bitnetNetwork()

BERT

MiniLM, BERT variants, LEAF/BGE/E5 embedders

sentence-transformers (MongoDB/mdbr-leaf embeddings)

bertNetwork()

gtr-t5-base (GTR embedder, vec2text inversion)

Real-weights gtr-base round-trip test

Hand-coded

Voxtral

Voxtral TTS

Not yet — architecture code only, no runtime facade

voxtralBackboneNetwork()

Documentation Structure

This documentation follows the Divio documentation system:

Tutorials: Step-by-step lessons to get you started.
How-to Guides: Practical recipes for specific tasks.
Reference: Technical descriptions of APIs and components.
Explanation: Background and design decisions.