BackContext is live: one context layer for every agent you run.
Explore BackContextThe weights are yours, so where the model runs should be your call. Here is how we deploy fine-tuned models inside your cloud account, what runs where, and how a new version gets to production without surprises.
The weights are yours in every case. You can move from managed inference to your own account later without retraining.
Our production models run 8B–32B parameters. An 8B model serves comfortably on a single modern data-center GPU. A 32B model usually needs one or two, depending on precision and traffic. Because these models are small and the work is repetitive, requests batch well, which is a large part of why cost lands at $1.40 per 1,000 tasks at production volume against $65.00 for GPT-6 Astra at list price.
Median latency for our production models is 0.4s, against 3.2s for Opus 5.5 and 5.9s for GPT-6 Astra on the same tasks. Inside your own network there is no public internet round trip, which helps. For workflows where a person waits on every answer, such as an order desk keying in phone orders, that difference is felt on every task.
Serving runs under your identity and access controls, in a network you define. Model weights and recorded data are encrypted at rest. API keys for fallback providers live in the BackContext vault, encrypted at rest with AES-256-GCM, and are released only per policy, with every request, approval and release written to the audit trail. Your security team reviews the same architecture document for every deployment, and we answer its questions before anything goes live.
Fallback calls go through your own provider account, so the same contracts and data terms you already have with Anthropic or OpenAI apply. If your policy does not allow any external calls for a workload, low-confidence tasks go to a human review queue instead. Either way, the fallback rate is on the dashboard. At DataFoldIT, it is 2.1% of tasks.
Requests and responses stay in your account. BackContext records each task and its outcome so the next version can learn from it, and that data is encrypted at rest. It trains your models and nothing else. It is never pooled with other customers and never used to train shared models.
That is often what gets a project through review. One customer's General Counsel put it simply: "The model runs in our VPC. Legal signed off in a week instead of a quarter."
A first model typically takes four to six weeks: two weeks of recording with BackContext, two weeks of training and evals, then a staged rollout behind the quality gate. The rollout has three steps.
New versions ship like code. Each one comes with an eval report against the current model and the four frontier models, passes the gate, rolls out in stages, and can be rolled back in one step. New verified outcomes keep arriving through BackContext, so the model gets better on your work over time.
Model weights you own, a model card, an eval report, and runbooks for serving, scaling and rollback. If we stopped working together tomorrow, the model would keep running. Read more about private deployment.
