Only as tooling during data preparation, never inside your production system, and only where terms permit. Anthropic explicitly permits using outputs to train models that don't compete with Claude, narrow tasks like classification, extraction and summarization are their own examples. OpenAI's terms prohibit training models that compete with OpenAI; a single-task classifier is generally understood not to, but we review each case. Google's Gemini API terms prohibit using the service to develop ML models at all, so we don't use it. When there is any doubt, or your policy forbids any external API, we rely on open-weights models only (DeepSeek MIT, Qwen Apache-2.0, Gemma 4 Apache-2.0, Mistral Apache-2.0, Llama community license) and the question disappears.
On open-ended tasks, yes, much worse, that's not what we sell. On a narrow task with a fixed output format, a compact model built for that task routinely matches or beats general frontier models, because it has seen thousands of examples of exactly your task and nothing else. The pilot measures this on your gold set before you commit.
An ordinary server. For most first tasks, classification, routing, extraction, a regular CPU machine you already have is enough. For very high volumes we may recommend adding one GPU card. We size it during the engagement and we don't sell hardware.
For the gold set: a few hundred examples checked by your expert. For training: a few thousand real inputs (unlabeled is fine, labeling is our job), or synthetic data generated from a description plus a handful of samples when real data can't be shared.
You retrain on the new base with the same dataset and recipe, a day of work, included in the care plan, or something your team can do with the pipeline we hand over. This is the advantage over an API: you upgrade when you choose.
Yes. Option B on the security page: data preparation and model training both inside your environment, with open-weights tooling only. We can work on-site or through a bastion you control.
No. Weights, dataset and pipeline are delivered to you and deleted on our side after acceptance, in writing. We may ask permission to describe the engagement anonymously as a case study; you can say no.
Anything the base models cover well. English, Russian, most European languages, Indonesian, Chinese. Russian-language tasks are a specialty; open bases with strong Russian (Qwen, GigaChat open weights, T-Pro) are available.
Yes. Engagements can be contracted through a local entity in rubles, all AI tooling is open-weights and hosted in-country, and deployment targets Russian clouds or your own servers. No dependency on foreign APIs in production.
Forward is built by an engineering team with a background in payment and fintech systems, environments where data residency and auditability were never optional. We are small, senior, and hands-on.
Two to four weeks. One task. A measurable result before you commit.