AiGanakAiGanak-SLM-1B
An in-house small language model built for private, long-document work inside a company's own infrastructure. It is still in training. This page tracks where it stands and what we have verified so far.
AiGanak-SLM-1B is AiGanak's in-house 1-billion-parameter decoder-only small language model with a 128k-token context window, built first for logistics and supply-chain documents. It is in Phase 2 convergence and has not been released.
Training progress
Benchmarks will be published when evaluation is complete.
What makes it different
The first use case is structured logistics and supply-chain data: manifests, tax IDs, coordinates and timestamps in messy real-world documents.
128k long-context stability
YaRN-scaled rotary positional embeddings with a 1,000,000 theta base, built to keep reasoning accurate across 131,072-token windows for large document retrieval and audit trails.
4-bit quantization-aware training
A custom straight-through estimator in the training loop prepares the model for near-lossless 4-bit precision on standard local hardware.
Structured attention steering
A steering mask prioritises structural tokens such as IDs, coordinates and timestamps, aimed at reducing hallucination on messy data.
Logistics-native vocabulary
Specialised BPE tokens for manifest and tax identifiers. After the vocabulary expansion the model re-converged within 400 steps.
Need your own model now? Start with a proof of concept.
Building SLM-1B gave us the training pipeline, data tooling and evaluation methods to build domain models for clients, in stages, on proven open or custom architectures.
- Enterprise data feed
- Local embeddings
- SLM-1B engine · 128k context
- On-premise inference
- Your application
Request early access
We are looking for partners where data privacy, air-gapped security and local-first AI are critical. Tell us about your data and hardware.
Questions people ask
Can't find it? Press ⌘K and ask.
What is AiGanak-SLM-1B?
An in-house 1-billion-parameter decoder-only transformer with a 128k-token context window, built first for structured logistics and supply-chain documents.
Can I use AiGanak-SLM-1B today?
No. It is in Phase 2 convergence and has not been released. Join the Registry for early access, or commission a custom model proof of concept.
What hardware is it trained on?
Training runs on AMD Instinct MI300X accelerators with ROCm and Flash Attention kernels.