Fine-Tune
Start from an open-weight base and train on prompts, completions, or conversation logs. Freuron runs versioned jobs on managed GPUs — typically within hours — so you ship a model fitted to your workload, not a shared generalist.
The infrastructure layer for production open-weight AI.
Turn production traffic into a dedicated open-weight model — trained, evaluated, and served on a private endpoint. Freuron keeps the loop running as new data arrives, so quality compounds without you owning the ML stack.
$freuron audit langfuse --window 30d --slug support-bot
›audit langfuse · 3 candidates ranked
›create support-bot · llama-3.1-8b-instruct
›importing support-2026.json · 12,481 qa
›train job_9b21 eta 1h47m
›epoch 3/3 train_loss 0.49
›evaluate eval_loss 0.41 quality gate passed
›deploy support-bot private endpoint
›POST /api/v1/inference/chat 200 · 184 tok
›monitor p95 184ms · feedback loop on
›retrain job_9c02 scheduled on traffic
$
audit → improveproduction open-weightprivate endpointslangfuse + helicone + otelopenai-familiar APIeval before switchauto-retrain on trafficper-tenant isolationtrain + serve, one bill
Fine-tune, deploy, and continuously improve open-weight models.
Freuron is the infrastructure underneath production open-weight systems — managed training, private inference, evaluation-gated cutovers, and traffic-driven retrain on serverless GPUs that scale to zero.
Deploy
Promote a validated version to a private, single-tenant endpoint via POST /api/v1/inference/chat. Cold start, scale-to-zero, and per-token billing are handled by the platform — no cluster to keep warm.
Continuously Improve
Production traffic and feedback feed the next training loop. Retrain on a schedule or threshold, evaluate the candidate, then switch the live endpoint — or roll back in one call. The MLOps loop stays on Freuron.
Production workloads deserve a dedicated open-weight model — not a shared generalist.
General-purpose endpoints are built for broad consumption and are typically over-provisioned for specific workloads. Freuron fine-tunes a smaller open-weight base on your data and serves it from a private, single-tenant endpoint — lower per-call cost, tighter latency for the target workload, and isolated customer data.
See which production workloads are ready for a dedicated model.
Connect Langfuse, Helicone, or OpenTelemetry in read-only mode. Freuron reviews cost, latency, and volume, then highlights the workloads that justify an open-weight fine-tune and private endpoint.
Production capabilities included with every model.
Private inference, evaluation-gated cutovers, persistent memory, web search, and traffic-driven retrain ship with the platform — not as separate tools you have to wire yourself.
Open-weight fine-tunes, no cluster to own
Submit JSON or JSONL prompt-completion datasets, including MongoDB Q→A imports. Freuron provisions compute, runs training on open-weight bases, and exposes a versioned model — without you managing GPUs or training scripts.
Private inference, production-shaped API
Deployed models land on serverless GPUs behind POST /api/v1/inference/chat. Same messages / choices shape as OpenAI chat completions — swap URL and key, nest generation params, and you are on Freuron.
Evaluate before the cutover
Training completes into a candidate version. Quality gates and eval loss surface before the live endpoint switches, so deploy is a deliberate promotion — not an accidental overwrite.
Persistent memory in the request path
Pass a user id and Freuron keeps cross-session context — preferences, tone, prior interactions — scoped per end user. No separate vector database to operate alongside inference.
Web search when the workload needs it
Toggle live web search per model. Results fold into replies with inline citations — useful grounding without a custom retrieval pipeline.
Monitor traffic, retrain on signal
New production data triggers automatic retraining. When a run passes evaluation, the platform can switch the live endpoint to the new version. Roll back to a prior version in one call when needed.
From your dataset to a live private endpoint in four API calls.
The path runs over HTTP. Freuron handles validation, training, evaluation-gated version switches, and serving — you do not provision or operate ML infrastructure.
Import training data
POST a JSON or JSONL file of prompts and completions. Freuron validates the schema, deduplicates rows, and stages the dataset for fine-tuning.
Fine-tune an open-weight base
Submit a fine-tune request on the model. The platform schedules training on managed GPUs, executes the run, and streams progress through the API.
Evaluate, then switch
When training completes, inspect the candidate and promote it when ready. Retrain on fresh production data later, or roll back the live model — each via a single POST. Freuron handles the active endpoint transition.
Deploy — call the private endpoint
POST /api/v1/inference/chat with an Inference key and the familiar messages payload. Response content lands at choices[0].message.content. The platform serves the request, scales to zero between calls, and bills per token.
One private endpoint. Familiar shape. Production-ready.
A single POST with an Inference key and the OpenAI-familiar messages / choices shape. Swap URL and key, nest generation params — memory, web search, and per-end-user identity are optional request fields.
access