Open source lab / models + agents

Open foundations for models and agents.

We build reusable primitives, developer tools, and hosted services for the hard problems shared across AI products.

Explore projectsView source
tokn-ai / open_stack
corpusffbpetokens
tokensflm.pymodel
requesttokn-routerinference
tracesllm-tokeiusage
open corecomposablehosted when useful

Showcase

Foundations in practice.

Representative outputs from the current public projects. Each one solves a different shared problem beneath AI products.

01of04

FFBPE training benchmark comparing regular and retained-bigram Unicode inventories.
tokenizationffbpe

Shape a multilingual inventory before training.

A measured release benchmark showing the effect of retained Unicode bigrams on one FineWeb2 Chinese workload.

View repository
FLM experiment configuration beside its composable package workflow.
trainingflm.py

Compose a complete model experiment.

Datasets, reusable modules, training, reinforcement learning, and inference remain visible parts of one workspace.

View repository
Tokn Router diagram connecting an OpenAI-compatible client to providers and local state.
inferencetokn-router

Route one local endpoint across providers.

An account-aware gateway handles compatible clients, provider selection, streaming, and local request state.

View repository
LLM Tokei sample usage table grouped by coding agent and model.
usagellm-tokei

Report where agent tokens went.

A representative grouped usage report across local coding-agent sources, models, token classes, and cost.

View repository

Open projects

Four projects, one direction.

Each project tackles a difficult shared layer. Use one on its own, or combine them without adopting an entire stack.

01tokenization

ffbpe

Fast and faithful byte-pair encoding for training exact tokenizers over large, multilingual corpora.

  • Python + Rust APIs
  • Exact bounded-memory training
  • Tiktoken-compatible encoding
availableRust / Python
02training

flm.py

A composable Python workspace for language-model research, training, reinforcement learning, and inference.

  • PyTorch foundations
  • Reusable model modules
  • Portable training experiments
research previewPython
03inference

tokn-router

One local, OpenAI-compatible endpoint for supported coding plans and model providers.

  • Single Rust binary
  • Streaming + account pools
  • Local SQLite usage records
in developmentRust
04usage

llm-tokei

Fast, local reports that show where your coding-agent tokens and estimated costs went.

  • Six agent sources
  • Human tables + stable JSON
  • No log uploads
availableRust

About the lab

The hard parts should be built once.

Most teams focus on the product they need to ship today. We focus on the foundations many products need, but few teams have the time to build well.

We build the tools behind AI products.

Open at the core. Local when preferred. Hosted when useful.
01_reusable

Useful beyond one product.

We invest in clear boundaries and durable interfaces so the same foundation can support many different ideas.

02_open

Understand every layer.

The core stays inspectable and replaceable. You should be able to see the tradeoffs and choose how each piece runs.

03_adaptable

Built for a moving field.

Focused components adapt better than heavy abstractions. We keep the pieces small enough to evolve and strong enough to depend on.

Rust / durable systemsTypeScript / product surfacesPython / model research

Build with us

Build on open foundations.

Explore the projects, use what fits, and help improve the layers beneath the next generation of AI products.

Explore GitHub Browse projects