Give a public-interest workload somewhere serious to run.

Crafter Research builds open, reproducible infrastructure for Latin America. We already have the corpora, benchmarks, pipelines, and public baselines. More compute would let us run controlled experiments that are currently too expensive, too slow, or impossible on rented API access alone.

Discuss a compute offer See shipped evidence

Three useful sizes. No all-or-nothing requirement.

We match the workload to the resource, publish measured utilization, and reduce scope before pretending a small grant can support a large study.

One bounded release

$1k–$5k cloud or API credits3–12 months

A fixed benchmark run, one open-model comparison, embeddings and reranking experiments, or a month of repeated public-system observations.

Public returnRaw outputs, exact configuration, cost ledger, limitations, and a Research Log publication.

Controlled model lab

300–1,000 A100-equivalent GPU-hours90-day allocation

Repeated serving of 7B–70B open models, quantization comparisons, retrieval ablations, and independent judging without API drift.

Public returnUtilization report, serving recipes, benchmark records, failure taxonomy, and reproducible evaluation code.

Persistent research node

High-memory workstation or equivalent remote capacity90-day loan or donation

Private pre-publication corpora, repeatable local inference, long-running observatory jobs, and experiments that do not fit into short cloud sessions.

Public returnMeasured workload report and a return, extend, or convert decision. Hardware never becomes the research claim.

What it unlocks

Named releases, not generic GPU hunger.

Each lane already has public inputs and a bounded next result. Scope is finalized against the actual resource offered.

LatamBench Open Weights Observatory

Which open models can Latin America actually run, and what changes under quantization?

Evidence ↗

Baseline: 11 models evaluated across Trueque-500 and CHOCLO-500, with published responses and judge records.
Compute job: Controlled inference for 6–8 open models, two quantizations, repeated runs, and independent judging.
Release: A comparable cost, quality, abstention, and serving-variance report with every response public.

Spanish Legal Model Reliability Lab

Can a system cite the law that was valid at the time of the question?

Evidence ↗

Baseline: 21,244 versioned legal norms and a firm N=35 retrieval gold set across six configurations.
Compute job: Expand to N≥150, compare RAG, long-context, and bounded local models, then repeat the strongest configurations.
Release: Per-query evidence, citation and temporal-validity metrics, failure taxonomy, and one-command verification.

Public Systems Observatory

Do public digital services remain reachable and usable after the landing page?

Evidence ↗

Baseline: Muniscan covers 1,871 municipalities and already publishes a deterministic 100-domain HTTP sample.
Compute job: Repeated network observations, browser-level accessibility checks, visual regression, and non-destructive citizen-task probes.
Release: Time-series evidence and a public failure taxonomy without turning one request into an uptime claim.

Good fit

What we can responsibly accept.

The best support has enough access time to complete a named release and permits publication of negative results.

Cloud: Credits valid for at least 90 days, with GPU or high-throughput model access.
Capacity: Reserved or interruptible accelerators with a documented hardware and software environment.
Hardware: High-memory NVIDIA or AMD workstations, server access, or measured evaluation loans.
Models: API credits for frontier controls, independent judges, embeddings, reranking, or multimodal agents.

Independence contract

Support funds the experiment. It does not purchase the conclusion.

01The provider and approximate value are disclosed.
02Methods, failures, utilization, and limitations remain publishable.
03Supported models receive no favorable scoring or editorial veto.
04Private data, credentials, and irreversible government actions stay out of scope.

Make the offer concrete

Send the resource, constraints, and expiry date.

We will reply with the best matching workload, a smaller fallback, milestones, stewardship requirements, and public deliverables. A useful email can be three lines long.

Email Railly Available resource / access window / restrictions railly@crafterstation.com ← All support options