Skip to content

A Soulstack product

Status: In developmentModel platform

Hollow Kiln

Kiln is Hollow’s model platform: repeatable local training, evaluation, inference, agent execution, experiment tracking, and traceable artefacts built around one set of canonical interfaces.

Build locally

A platform for training and evaluating models

Kiln keeps model work local and inspectable, from tokenizer and dataset through training run, checkpoint, evaluation, inference, and agent execution.

  • Repeatable training and evaluation workflows
  • Traceable datasets, checkpoints, and model lineage
  • A FastAPI backend, CLI, and React UI over the same interfaces
KilnThe platform
  • 01Training and evaluation
  • 02Local inference
  • 03Experiment tracking
  • 04CLI and web UI

Model families

Three directions, one careful platform

Kiln is developing Kindling, Ember, and Soot as distinct model families. The family pages explain the research direction without presenting checkpoints or pilots as finished releases.

  • Kindling, Ember, and Soot have separate roles and lineages
  • A trained checkpoint is not automatically a public release
  • Weights, data policy, licence, and limitations stay explicit
FamiliesThe work in progress
  • 01Kindling: small and efficient
  • 02Ember: primary family
  • 03Soot: open family
  • 04Release status stays explicit

What’s inside

Everything around the model matters

  • TrainingLocal, repeatable runs with checkpoints and recovery.
  • EvaluationCapability measurement and regression tracking against recorded baselines.
  • InferenceModel loading, generation, serving, and quantisation paths.
  • AgentsTool-using model runs with explicit permissions and recorded context.
  • LineageModels carry the tokenizer, dataset, configuration, and run history that produced them.
  • Local firstThe platform is designed around local experiments and inspectable artefacts.

Build models with a complete record of how they came to be

Hollow Kiln is in active development. Its model families are documented as research work, not public downloads.