UnderstandSmall language models (SLMs)

Own the model. Don't rent it.

A small model trained for your task often matches a large general model on that task. It runs on your servers or in your cloud account, your data stays in your environment, and the trained weights are yours under the engagement terms; the base model's license still applies.

Start with a Decision BlueprintTalk to the engineer

FIG. 1 · Your server · one base modelIllustrativeTask by task

An adapter trained for each task.

Open-weight base model, loaded once · its license still applies · no outside model service

Statement spreading
Supervised fine-tuning (SFT), as a low-rank adaptation (LoRA) adapterPassed
Covenant check
SFT, then reinforcement learning (RL)Passed
Complaint routing
SFT, as a LoRA adapterPassed
Group structure
From scanned annexes; the test is still runningNot yet

Your expert writes the test. The model goes live only on the tasks it passes.

When it fits, and when you don't need us

A trained small model fits where a general one falls short.

  • Your experts read the same documents every day

    A general model misreads your terms and your house format, and at volume every page sent to a paid model adds to the bill.

  • Your data may not leave

    Customer files, memos, and recorded calls stay on your servers; the model comes to the data.

  • You don't need us

    A template or a general assistant already does it

    If it passes the test your expert writes and your data may leave, use it.

Not sure which kind fits your decision? The Decision Blueprint answers that first.

How we build it

Post-training: from a general model to your expert.

Post-training is further training on examples your own experts have already decided; it makes a small model your expert on one task. Before it starts we build the right data and practice tasks; then each step runs only where your expert's test shows it is needed.

For your technical team: what we prepare before training

Before training: the right data and environments

  • A test set your expert grades
  • Worked examples from your decided files (SFT data)
  • Expert-reviewed synthetic data
  • Practice tasks with answers your team can check (RL environments)
  • Personal data removed
  • Data sources, traced

FIG. 2 · Post-training · credit memo modelIllustrative

  1. Base model

    Knows language, not your desk.

  2. Continued pre-training (CPT)

    If needed

    Only if your field's language is missing.

  3. Supervised fine-tuning (SFT)

    Core step

    Your formats, from your decided cases.

  4. Reinforcement learning (RL)

    If a reviewable reward exists

    Practice against a reward your team can review.

  5. Your expert's test

    Graded by your expert before training; held back from it.

One task: a credit memo's covenant section

Base model

“The company appears stable overall. Some covenants may be under pressure.”

  • Generic wording
  • Not your section format
  • No source

Your data, weights, and test set stay with you and never train another system.

Your expert writes the test. The model goes live only on the tasks it passes.

What it can be trained for

  • Domain-specific

    Your field's language: credit, claims, maintenance.

  • Task-specific

    One job: a statement spread, a covenant check, a summary.

  • Institution-specific

    Your formats, policies, and terms.

  • New expertise

    A skill the base model never had.

  • Tools and memory

    Calls the right tool with the right parameters, such as a SQL query or a JSON request, and keeps what the case needs across steps; the agent's harness decides which tools it may use.

  • Text, images, speech

    Scans and defect photos through vision-language models; your calls through speech recognition and synthesis.

In visual inspection a model trained on good parts makes the pass-or-fail call; the vision-language model names the defect. No speaker identification or voiceprints. No video.

Facts that change, such as rates and product terms, are read from your sources at answer time through retrieval, with the source cited. Policy and authority stay outside the model weights.

Rent or own

Renting a general model, or owning your own: the difference.

A rented general model (API)A small model you own
Where data goesEvery request leaves your network for the providerStays on your servers or in your cloud account
Who holds the weightsThe providerYou, under the engagement terms; the base model's license still applies
Change controlThe provider can update or retire the modelChanges only when your team approves a version; the way back stays with you
LeavingPrompts move; behavior may notWeights, data, the training recipe, and the test set are handed over
Cost shapePaid per call; the bill grows with volumeSized hardware; at volume often cheaper
On your taskBroad knowledge; may misread your terms and your formatTrained on your decided files; on that task often matches a large general model
ProofGeneral benchmarksYour expert's test, task by task
Where it fits betterLow volume, open-ended questions, data that may leaveRepeated, narrow, regulated work

We size the hardware in the Decision Blueprint, measured on your own cases.

One base model, an adapter per agent

Each agent with its own expertise, on one base model.

The base model is loaded on your server once. A small adapter is trained on top of it for each task or agent, and an incoming file runs only its own adapter.

Incoming file: a loan 12 days lateIllustrative

  1. AgentCredit agentAdapterStatements and the credit memoSFT · LoRA
  2. AgentCollections agentAdapterPayment plan draftingSFT · LoRA
  3. AgentComplaints agentAdapterReplies in one client's formatSFT, then RL

Open-weight base modelloaded once · its license still applies

Your server · your environmentNo outside model service

  1. Supervised fine-tuning (SFT), with LoRAThe core step: your format and your task, from files your team already decided.
  2. Reinforcement learning (RL)Only where there is a reward your team can review.
  3. Continued pre-training (CPT)Only where your field's language is missing from the base model.

A new task gets a new adapter, not a new model. We size the hardware in the Decision Blueprint.

What you receive

Weights that are yours, running where your data is.

FIG. 4 · HandoverIllustrative

  • model/The weights, yours under the engagement terms; base-model licenses still apply
  • data/The datasets, the training recipe, and their sources
  • test-set/Your expert's graded test set
  • evaluation-reportTask by task, against that test
  • serving/Serving in your environment, with monitoring
  • rollback/The path back to the previous version

Where it runs

On your own servers, cut off from outside networks where required, or in your cloud account.

A small model trained on your task often matches a large general model on that task, and often runs on a single server in your own data center. We size the hardware in the Decision Blueprint.

Training a model from scratch is not part of the work unless it is separately scoped.

How your data, model weights, and access are handled: Security & working with us

Desks that use it

Desks where a small model reads, reasons, and calls your tools.

  • SLMs

    Corporate credit memo and financial statement analysis

    TodayAnalysts spread statements and write the memo by hand, in a fixed format, for every renewal.

    With the systemStatements spread with their sources, and the figures to check flagged.

    Credit analyst decides

  • SLMsAgentic AI

    Regulation to policy

    TodayA new circular arrives, and nobody knows which sections of your policies it changes.

    With the systemEach policy section marked changed or not, with the reason on record.

    Compliance officer decides

  • SLMs

    Request and complaint routing

    TodayThousands of messages a day land in one queue; misroutes miss legal deadlines.

    With the systemEach message routed to the desk that owns it; the model never closes one.

    Head of operations decides

  • ModelsSLMsAgentic AI

    Early arrears collections

    TodayCalls reach customers who would have paid anyway, while the ones who cannot keep up need a plan, not ten calls.

    With the systemHardship told apart from won't-pay, with the plan's arithmetic on file.

    Collector decides

  • ModelsSLMsAgentic AI

    Visual inspection

    TodayInspectors check every part under a lamp; night shift misses more, and each nonconformance report takes half a day.

    With the systemA draft nonconformance report with its evidence, ready for a hold-or-release call.

    Quality engineer decides

  • SLMsAgentic AI

    Contact-center speech: quality and compliance on every call

    TodayEvery call is recorded, quality hears two in a hundred, and off-the-shelf speech-to-text mangles your product names.

    With the systemEvery call checked for quality and compliance; the coaching point in context.

    Team lead decides

See all 6 desks

It works on its own, and stacks with:

Field note: Retrieval or fine-tuning: ask about the data first

Equalize the gap.

Bring one document task. We test it on your files first.

Tell us the documents, the task, and how a good answer is graded today. The Decision Blueprint shows whether a small model can pass your test.

Start with a Decision BlueprintTalk to the engineer

https://eqlx.ai/small-language-models · Printed copy; the website is the current version.