BONACCI FOUNDRY · MODEL SERVICES

The model runs on your GPUs.
We build the model.

Domain models trained on your data, inside your network, delivered as weights you own outright. Fine-tuning where that is enough. Continued pretraining where it is not. Pretraining from scratch when your data has no usable base model.

One week scoping engagement first · full builds $15k to $120k · trained in your environment or ours

Why this exists

Private inference on a model that has never seen your data

Studio already runs inference on your own hardware. No prompts leave, no schemas leave, no sample rows leave. That solves the sovereignty problem.

It does not solve the quality problem. The checkpoint you point Studio at is a general open-weights model trained on public web text. It has never seen your product taxonomy, your regulatory vocabulary, your customers' languages, or your internal document formats. You get privacy and mediocrity in the same package.

Foundry closes that. Same perimeter, same GPUs, a model that actually knows your domain.

vllm.internal.acme-bank.corpsame box
general checkpoint
> classify: "NOSTRO recon break, val dt T+2"
label: payment_issue
confidence: 0.41
tokens in: 31
private. does not know your vocabulary.
your domain model
> classify: "NOSTRO recon break, val dt T+2"
label: nostro_break.value_date
confidence: 0.96
tokens in: 14
private. trained on your ground truth.
Honest scoping

Four situations. One of them is "don't do this."

A custom model is the right answer less often than the market suggests. These are the cases where it holds up, and the case where it does not.

Your vocabulary breaks general models

Domain jargon, regulatory language, internal taxonomies, and document formats that public web text never covered. Accuracy on your task, not on a benchmark.

Inference cost is a line item

A general frontier API on a repeated narrow task is the most expensive way to solve it. A small model trained on that one task runs at a fraction of the cost, on hardware you already own.

The data cannot leave the building

Training data is the most sensitive data you have. If a hosted fine-tuning API is closed to you for the same reasons hosted inference is, this is the only remaining path.

You probably don't need this

Most teams asking for a custom model need retrieval. We will tell you that in four days and charge you for the four days, not the four months. The projects that survive that filter are the ones worth doing.

What we build

Most people who ask for a model from scratch should not buy one

There are three rungs. Picking the right one is most of the value we add, and it usually saves you money.

01

Fine-tuning

You want new behaviour, format, or task performance.

Instruction tuning or LoRA on an open-weights base. Delivered as safetensors and GGUF, deployable on your existing vLLM or Ollama setup.

2 weeks
from $15,000
02

Continued pretraining

You want new domain knowledge, not just new behaviour.

Corpus construction from your own data, continued pretraining on an open base, then post-training. Delivers most of the value of a from-scratch build at a fraction of the cost. This is the right answer for the large majority of teams who arrive asking for rung 3.

3 to 5 weeks
from $30,000
03

Pretraining from scratch

Only correct when one of these is true.
  • Your languages tokenize badly on existing vocabularies, making general models both expensive and inaccurate
  • Your data is not natural text: sequences, telemetry, logs, structured signals with no usable pretrained base
  • Every training token must be certifiable, so inheriting a model with an unauditable corpus is not permitted

Tokenizer design, corpus pipeline, architecture selection, pretraining, post-training, eval harness, technical report.

6 to 9 weeks
from $60,000
compute billed at cost plus 15%
Tell us which rung you need Book a scoping call

The clock starts at data access, not at signature. The training run is rarely the long pole: a 9 billion parameter model over 122 billion tokens is days of wall clock on the right pod, so we compress by widening the pod and passing the compute through at cost, not by cutting evaluation. Corpus readiness is what actually moves the date.

Process

Scoped before it is sold

Every engagement starts with a paid audit that can conclude you should not build a model at all.

STEP 1

Scoping engagement

One week, priced separately. We audit your data, assess tokenizer fit, estimate a token budget, and recommend a rung. You get a written report whether or not you continue. Roughly half of these conclude that retrieval or a fine-tune is enough, and we say so.

STEP 2

Corpus construction

Your training data is built as a versioned pipeline with contracts enforced at compile time: PII redaction verified, class balance checked, language mix asserted. Lineage is emitted, so you can answer which source rows produced which model version.

STEP 3

Training and evaluation

Runs in your environment, your cloud account, or an India-resident GPU provider, whichever your compliance boundary requires. Evaluation is against a held-out set built from your ground truth, not public benchmarks.

STEP 4

Delivery

Weights you own outright, in safetensors and GGUF. Eval harness and training configuration handed over. Deployable behind Studio on your own vLLM or Ollama, or standalone.

Billing is by milestone: 30% on corpus delivery, 30% at the mid-training checkpoint with evaluation, 40% on final delivery. Compute is passed through at cost plus 15% and billed separately.

Depth

We have been to the bottom of this stack

Most vendors offering custom models are running someone else's fine-tuning API behind a project plan. We pretrained a 9 billion parameter mixture-of-experts model from scratch: 122 billion tokens, Google TPU v6e Trillium, MaxText and JAX. Tokenizer, corpus pipeline, architecture, training loop, evaluation.

We published the full technical report, including the benchmarks where it underperforms and the decisions we would make differently. That is the level of detail you should demand from anyone you let near your training data.

Read the technical report →

9B
parameters pretrained from scratch
122B
training tokens
TPU v6e
MaxText and JAX
Open
published report and eval harness
Ask first

Not ready for a call? Send us the details.

A model engineer reads every one of these and replies in writing, usually within two working days. No sequence, no nurture track.

  • 1
    The task you want the model to do, in one or two sentences.
  • 2
    Roughly what data you hold for it: volume, languages, and whether it is natural text.
  • 3
    The constraint that rules out a hosted API, if there is one.

Prefer email? Write to contact@thinkingdbx.com. If your question is really "which rung do I need", a scoping call answers it faster.

Start here

Start with the scoping call

Bring the task, the data you have, and the constraint you are working under. We will tell you which rung you need, including when the answer is none of them.

Or email contact@thinkingdbx.com.

  • 1

    Weights

    safetensors and GGUF, owned outright, no licence to renew.

  • 2

    Eval harness

    Built on your ground truth, so you can re-run it on every future version.

  • 3

    Corpus pipeline

    Versioned, contract enforced, with lineage from source rows to model version.

  • 4

    Training configuration

    The exact configs and report, so another team could reproduce the run.