TinyMachines Labs · Post-training research

Small models,trained for
real work.
Built for each vertical.

TinyMachines Labs is a research lab post-training small models to state-of-the-art performance in specific verticals, using full-parameter fine-tuning, distillation and reinforcement learning.

01

The stack

How a model moves through our stack: data, full-parameter SFT, then full-parameter RL with verifiable rewards, held-out evaluation, and finally our cloud harness, where trained models are put to work and tested.

The TinyMachines stackInputs feed a data pipeline and registry, then full-parameter SFT produces a cold-start checkpoint. An RL environment used only in training generates scored episodes for full-parameter GRPO, which produces the specialized checkpoint. That checkpoint is evaluated on held-out tasks and then served, frozen, in the cloud harness used for testing and demos, not for training.L0InputsExternal datasetsSeed datasynthetic · teacher · solverTask definitions→ RL environmentsBase modele.g. Qwen3.5-4B → SFTL1Data & registryDownloaderlicense gateIngestadaptersTransform chainverifier gate · dedupquality · PII scrubloss mask · packSplittrain · val · held-outRegistrydata · checkpoints · runsL2Full-parameter SFTSFT trainerfull-parameter · packingCold-startcheckpointL3RL environmenttraining onlyvLLM generatorcurrent policyTraining harnesstyped toolsTask envsno-egress sandboxVerifiersrewardsEpisode recordstrajectories · rewardsRL replayL4Full-parameter GRPOGRPO trainerKL to frozen reference · curriculumSpecializedcheckpointOrchestratorrollout groupsweight syncL5EvaluationEvalpass@k · held-out transferReportsto the registryL6Cloud harnesstest & demo, not trainingUser sessiontest · demoCloud harnessorchestrator · specialistsModel servervLLM · frozen weightsSandboxno-egress · verifier gatestool callsloadedfrozen
02

Why small models

Pre-training sets what a model could know. Post-training settles what it reliably does.

Much everyday AI work is narrow and repetitive: the same kind of request, thousands of times a day. A large general model can do it, but often a much smaller one could do it as well once it has been trained for that task.

Small models are cheaper to run, faster to answer and easier to host privately or on a device. Our research asks when and how they can close the gap with frontier models, and how to tell, honestly, whether they have. Once a model is trained, we run it on our own cloud harness to test it on the real work it was built for.

03

How we work

Every model we train goes through the same four steps. Each one is versioned and reproducible: tasks, evaluations and training recipes are recorded with every run.

01

Define the task

Write down the exact task and what a good result looks like, as a versioned spec, before any training starts.

Task spec
Success criteria
02

Build the evaluation

Turn real examples into held-out tests, so every change is judged on the work itself rather than on generic benchmarks. Evals are versioned and rerunnable.

Held-out sets
Rubrics
03

Post-train

Full-parameter fine-tuning, distillation from larger teachers and reinforcement learning on verifiable tasks, run as reproducible recipes, starting with the lightest method that could work.

SFT · Distillation
RL
04

Ship only when it wins

A model is promoted, or a result written up, only when it beats the baseline on quality, cost and speed on the same evaluation.

Quality · Cost
Latency
04

Who we work with

We learn most from real problems. We'd like to collaborate with, and learn from:

Teams running repetitive work on frontier models

When the same kind of task runs thousands of times a day, it is a good candidate for a smaller specialist, and a good research problem.

Products that need low latency or on-device models

Tight response budgets and limited hardware are constraints we design around, not work around.

Organisations that need private, self-hosted weights

Where data cannot leave your environment, small models you can host yourself are worth studying carefully.

Researchers are welcome too. We work with each collaborator on one well-defined task at a time.

05

Research notes

Our first research note is published. More notes, papers and releases will be listed here as they appear.

06

Principles

iMeasure first

A result is only as strong as the evaluation behind it. We publish both, limitations included.

iiReward the outcome

Where a task allows it, RL rewards come from programmatic checks on the result, not from a model's opinion of it.

iiiDesign for the budget

Compute, memory and latency are constraints we design around. Efficient models reach more people.

07 · Pilot

Start a 6-week pilot on one real task.

In six weeks we take one well-defined task you run at scale, post-train a small specialist model for it, and measure it against the model you use today on the same held-out evaluation.

Scope

  • Weeks 1–2: define the task and success criteria, and build a held-out evaluation from your real examples.
  • Weeks 3–5: full-parameter fine-tuning, distillation and reinforcement learning on a small base model.
  • Week 6: test the trained model on the real work and compare it with your current model.

Deliverables

  • A versioned task spec and success criteria.
  • A held-out evaluation set you keep and can rerun.
  • A trained specialist model for the task.
  • A written report on quality, cost and latency against your baseline.

Who it’s for

  • Teams running one narrow, repetitive task thousands of times a day on a frontier model.
  • Products that need lower latency, lower cost or private hosting.
  • Teams that can share real examples and say what a good result looks like.
Start a 6-week pilot →or write to hello@tinymachineslabs.com