Language models deployed inside your network, with the documentation an audit asks for.
We deploy and fine-tune models on infrastructure you control, and we produce the evidence trail that follows: model provenance, evaluation records with stated methodology, access and retention controls, and reproducible build details for the hardware. The output is a system someone else can inspect and a file someone else can defend.
This is written for organisations that cannot send a corpus to a third party, and for those that expect to be asked to account for how a model was chosen, what it was trained on, how it was measured, and who can reach it.
Two ways to start. Most projects need both.
The data cannot be sent to a third-party API.
Curation, prompt formatting, LoRA or full-weight training, and an eval set written before training starts, so “is it better” has an answer that does not rest on our opinion. The training runs on your hardware, or on ours under an agreement you read before anything moves.
Talk about a fine-tune Machine workThe model needs a machine to run on.
A server sized to the model and the query load, not to a wish list. We spec, source and build it, burn it in under sustained load, and hand it over with the parts list, the thermal readings and a build guide somebody else could follow. Air-gapped if the work requires it.
See the reference builds1 — Evaluate
We take a slice of your data, run one existing model against it, and give you a measured answer: whether a fine-tune is viable, roughly what it would score, and what it would cost to run. An evaluation report with the numbers in it, or a plain no and the reason for it.
2 — Fine-tune
Curation, training and evaluation on your corpus. You keep the checkpoint, the pipeline and the harness. The price is fixed before the work starts, so the number in the brief is the number on the invoice.
3 — Model and machine
The fine-tune plus the server it runs on: specced, built, burned in and documented down to part numbers, handed over with the model already serving behind your auth. Parts at cost, the build fee on its own line.
Beyond those three we quote to the job. Every price above is fixed before work starts, and none of them improves if you decide this week.
What the engagement includes.
Usually all of it in one project. You cannot evaluate what has not been trained, and you cannot hand over what has not been evaluated. The deliverable is not only the model and the machine: each engagement closes with the written record that supports your own assessment of the system afterward.
Fine-tuning on a corpus that cannot move
LoRA, QLoRA or full-weight, with the eval set written first and the baseline recorded before a single step runs. For a house dialect, an internal shorthand, or a corpus with too little published text for a general model to have absorbed. About 8B parameters is the practical ceiling for full-weight on one 96 GB card; QLoRA on a 70B base fits with room to spare.
Deployment on your own network
vLLM, TensorRT-LLM or GGUF on a machine you control, behind your authentication and your logging. Nothing in the inference path calls a third-party API, which you can verify by observing network traffic rather than taking our word for it. Air-gapped on request, with the access and retention controls written down.
Server builds, documented
Sized to the model and the query load rather than to a spec sheet. You get the parts list with part numbers, thermal numbers under sustained load, the software bill of materials, and a build guide detailed enough to hand to another supplier.
Evaluation records with a stated method
A fine-tune without a harness is a claim. You get the eval set, the baseline scores, the exact package versions and the methodology behind them, so a later change can be shown to be an improvement rather than argued to be one, to a board, an auditor or an IRB.
For this work, a cloud
API is not an option.
Export-controlled drawings. Fab process logs under an NDA. Patient records. Privileged documents, where handing them to a third party waives the privilege. Wafer yield data that is the entire competitive position of the company holding it.
For that work the question is not which model tops a leaderboard. It is whether the weights, the corpus and the inference stay on a machine you control. That is the whole of what we build.
It is a narrower promise than it sounds, so here is the edge of it. It covers the inference path and the storage. It does not cover a clinician emailing a file from a phone, and no machine we build will stop that. Where it does apply, you can audit it rather than believe it: an air-gapped build keeps serving with the network cable out, because nothing upstream is required.
How deployment works →The corpus stays where you put it
Training happens on your hardware, or on ours under an agreement you read before anything moves: which machine, who can reach it, what is deleted at the end and when. That agreement is the part people actually audit, so it gets written first.
You keep the weights
Safetensors, GGUF or ONNX, serveable without us and not licensed back to you. The training pipeline and the harness come with them, so next year’s corpus runs through the same method and is compared against this year’s scores.
Air-gapped, if that is what it takes
A machine that never touches a network, built and burned in before it ships, handed over with the model already on it and the documentation an auditor will ask for. It is slower to support afterwards, and that is the trade.
The work these machines are for.
If your work resembles one of these, the design already exists and something close to it has been run. If it resembles none of them, describe it anyway; the unusual cases are usually the ones where this is worth discussing.
Fine-tuning
QLoRA on a 70B base fits one 96 GB card with room to spare, and it will go past 100B at 4-bit. Full-weight, every parameter trainable, is a different budget: about 8B on the same card. Long runs are fine because checkpoints survive between sessions.
Inference & serving
vLLM and TensorRT-LLM with continuous batching. 96 GB holds a 70B model at 4-bit with a long context, which is the awkward case and the one most on-premise deployments end at.
Simulation
GROMACS, OpenMM and similar molecular dynamics work, plus CFD, N-body and FEA. Multi-day runs are normal here, and a machine you own has no scheduler that reclaims it at midnight.
Agents & long context
SPARK-1’s 128 GB unified pool suits multi-step agent workflows and RAG pipelines, and fine-tunes whose working set is larger than a card’s VRAM.
Diffusion & vision
Training and sampling for SD3, FLUX.1 and similar, at the batch sizes smaller cards make you juggle memory for.
Domain translation
A house dialect, an internal shorthand, a language with too little published text. Narrow corpora are where a small fine-tune beats a large general model, and where owning the hardware pays for itself.
The machines we recommend
are the machines we use.
It is the only honest basis for a hardware recommendation: whoever gives it has run the job on the same parts. Here is what is in the rack, down to the part numbers.
Full specifications →The bill of materials for the machines we build, at the part numbers we actually buy.
What we built for our own work.
Each of these started as a tool we needed. That is why we can tell you where they break, and the same reason we can tell you whether a machine suits a job: something of ours has already run on it.
ToxScreen
Predicts hERG cardiac blockade and cytochrome P450 liability from a SMILES string in seconds, with a calibrated confidence score on every prediction rather than a bare label. Built on the published ADMET-AI architecture with uncertainty quantification added on top. Free to try, no account.
Lithic
Describe a part in plain English and get CAD geometry, toolpaths, G-code and a manufacturing report back, with cycle time and material use estimated. It exists because we needed our own fixture work turned around faster, which is how most useful internal tools start.
Naxi fine-tuning
A Qwen3-14B run on 62,021 weighted examples of Naxi, a language with almost no published text. Measured on a 446-row held-out gate it returns 45.78 chrF++ and 69.07 NTQS. Same method, applied to whatever corpus you have.
What you can check before you call.
Ours, and public
Built and run by us. The weights are public, the rubric is public, and scoring is done by script rather than by judgment.
Naxi language model
A Qwen3-14B fine-tune on 62,021 weighted examples of Naxi, a language with very little published text. On a 446-row held-out gate, scored with paired bootstrap confidence intervals clustered by reference, it returns 45.78 chrF++ and 69.07 NTQS: two independent metrics, reported together because one alone cannot separate a good sentence from a well-formed one.
NTQS is our own rubric: five weighted components, scored entirely locally, no model acting as judge. Weights, rubric and scoring script are on Hugging Face, so the number can be reproduced instead of taken on trust. It is also the case that matters for a narrow corpus: a small model trained on your data beats a large general one, and you keep it.
ToxScreen
Predicts hERG and cytochrome P450 liability from a SMILES string, with a calibrated confidence score on every prediction instead of a bare label. Built on the published ADMET-AI work, with uncertainty quantification added on top.
What it replaces for a screening group is a queue on a shared cluster: run the whole library on a machine you own and get uncertainty on every row, when you want it.
Lithic
Describe a part in plain English and get CAD geometry, toolpaths, G-code and a manufacturing report back, with cycle time and material use estimated. Built on the DeepCAD generative approach.
It exists because we needed to turn our own fixture work around faster. That is how most useful internal tools start, and it is the reason we can tell you where it breaks.
How the work usually arrives
These are the shapes projects arrive in rather than client references. On a call we will tell you what we have built for other people, which parts we can show, and where our experience stops.
A model on data that cannot move
Decades of operational records, commercially sensitive and contractually immovable. The work is curation, training and — the part people skip — an evaluation set built before training, so “is it better” has an answer that does not depend on our opinion.
You end up with the checkpoint, the pipeline, the harness and the numbers, and can re-run the whole thing without us.
A machine instead of a subscription
When a team works out it would keep a node busy most of the month, the cloud stops being the cheap option. We specify against your actual workload, source, build, burn in, and hand over the parts list, the thermal numbers under sustained load and a build guide.
Nothing in the machine depends on us afterwards. That is the point of buying it.
Making an answer defensible
The hard part is rarely the training run. It is producing something a board, an auditor or an IRB will accept. We ship the eval set, the baseline numbers and the exact package versions, so a later change can be shown to be an improvement rather than argued to be one.
One 96 GB card, three ways to end up with it.
Only the first number belongs to somebody else; the third is ours and the second is whatever your parts cost. The point is not that owning is always cheaper — it is that you should be able to see the three of them side by side and decide in a minute instead of a month.
Rented, same card
Runpod’s published on-demand rate for an RTX Pro 6000 with 96 GB, read off their pricing page on 31 October 2026. Held around the clock that is 8,760 hours, or about $21,800 a year, and at the end of it the hardware is still theirs.
You get elasticity and somebody else’s operations. You do not get the machine, a price that cannot move, or a stack on your own network where you can watch the traffic.
Own it, assembled yourself
A 96 GB Blackwell card, a platform to hold it, storage that will not throttle under sustained load, and a power supply sized for the transient rather than the average. You save the build fee.
You also own the integration, the burn-in, the driver versions, the part that arrives dead, and the afternoon it will not post. A fair trade for some teams and a bad one for most.
Built by us
The card and platform specced against your model, assembled, burned in under load, documented down to part numbers, and handed over with the fine-tune already serving behind your auth.
Parts at cost, the build fee on its own line, and a machine any competent shop can repair from the parts list if you would rather not call us.
If what you actually need is a record of patient information leaving a practice, that is a different product at a different price. Gatechart runs $299 to $849 a month; the incumbent HIPAA software it sits beside is $169 to $679 a month at Accountable’s published annual rate, and produces documents you fill in. Gatechart does a narrow job, and that page is explicit about the six things it does not do.
What happens after you send it.
Four steps, and the first one costs you a paragraph. If the honest answer is that renting would suit you better, that is the answer you get on the call rather than after an invoice.
- Where
- Manhattan, New York
- Phone
- (917) 667-1002
- First reply
- Two working days
- Prices
- $2,500 – $25,000
You send the brief, or you call
Whichever is faster. The form goes to info@catalyticgpu.com, which the people who build the machine read. Two working days is the honest worst case; it is usually the same day.
A thirty-minute call, no deck
What the model has to do, where the corpus sits, and who has to approve the spend. If retrieval would do the job instead of a fine-tune, that is what you hear on this call.
A fixed-scope pilot, invoiced up front
Two weeks. $2,500 to evaluate, $7,500 to fine-tune. If the corpus cannot support the outcome you want, you get that in writing in week one and we stop — what you bought is the answer, not the attempt.
Handover, then nothing that depends on us
The checkpoint, the harness, the numbers, and if you bought one, the machine with its parts list and build guide. Support afterwards is optional and priced separately. If you never call again, it keeps working.
Eight questions, answered as we answer them on a call.
These come up every time, including the ones with an answer you will not like. If yours is not here, ask it directly: (917) 667-1002.
Who are you? We have not heard of you.
Catalytic AI is a small firm in Manhattan: one rack, the people who built it, and no account manager in between. The phone number on this page reaches them. A firm this size carries a real risk, so the handover is built for it: if we disappear you keep the weights, the harness, the parts list and the build guide.
How do we verify that the data really does not leave?
Do not take it on faith; instrument it. Which machine holds the corpus, who can reach it and what is deleted at the end is agreed in writing before anything moves. Then check it yourself: an air-gapped build keeps serving with the network cable out, because nothing upstream is required, and a networked build can be packet-captured for a week and the capture read.
What happens if the machine breaks?
The parts list has part numbers, so any competent shop can source a replacement, and the build guide is written for exactly that. Support is optional and priced separately; when you take it you are talking to whoever assembled the machine. A remote session is arranged with you, time-boxed, and revoked at the end.
How long does it take?
A fixed-scope evaluation is two weeks. A full engagement, including a machine, runs two to eight weeks depending on the corpus and on parts lead times. If it overruns you hear it in the week we know, not the week it was due.
Is this expensive?
$2,500 to evaluate, $7,500 to fine-tune, from $25,000 for model and machine. All three numbers are printed further up this page, which is unusual in a category where most vendors make you book a call to hear one. The fixed prices are fixed: the number in the brief is the number on the invoice.
Is renting a GPU cheaper?
Sometimes, yes, and we will say so. Two hours a week, rent. Busy most of the month and the rent stops being cheap: the same 96 GB card is $2.49 an hour on Runpod’s published list, about $21,800 a year run continuously, and it is still their card at the end. We do that arithmetic with your numbers, including the version where it says rent.
Can our own team do this?
Probably, with enough weeks. What you are buying is not the ability to run a training script. It is the eval set written before training, the baseline numbers, the burn-in under sustained load, and the documentation that survives an audit. If your team already has those, you do not need this page.
Does this make us compliant?
No, and no software does: a scanner does not certify an organisation. We build and document the machine and the model; whether that is sufficient is decided by a regulator, an auditor or your counsel. What we produce is the record they will ask for, designed to be mapped to whichever framework applies to you — the NIST AI Risk Management Framework, a state AI statute, a sector rule, a procurement questionnaire or an internal audit — and we do not represent any of it as legal compliance. If what you need is a record of patient information leaving a practice, that is Gatechart, $299 to $849 a month, and its page says plainly what it does not do.
Send the brief. Two working days to a real answer.
A paragraph about what the model has to do and what is in the way is enough to start. If a fine-tune is the wrong tool for it, you will hear that, and you will hear it before an invoice.
CATALYTIC AI · MANHATTAN, NEW YORK EVALUATION $2,500 · FINE-TUNE $7,500 · MODEL AND MACHINE FROM $25,000