Local decision models: decisions, not text
We train and deploy on your premises small specialized LLMs that classify, route and score your business decisions — with calibrated probabilities, never generating text. 4 GB of VRAM is enough.
An LLM that decides
The model reads your input once, evaluates every question and option, then returns a decision with its probability. No text generation: deterministic answers, low latency, minimal memory footprint.
Calibrated accuracy
Up to 81% accuracy on public decision benchmarks (decontaminated test set), with calibrated probabilities to define your confidence thresholds.
Ultra-lightweight
Compact models (0.8B to 4B parameters): 4 to 15 GB of VRAM is enough. Runs on a single GPU, even consumer-grade.
Offline-capable
On-premise or full air-gap: no data leaves your network, neither during training nor inference.
Three decision types
Each decision is defined by your business criteria. The model returns the answer and its probability for every question asked on a single input.
Choice
Selects an option from your criteria: team routing, case category, recommended action.
Which team should handle this ticket? → "billing" (p = 0.94)
Yes / No
Answers a binary question: refund request, compliance alert, unusual clause detected.
Does the customer ask for a refund? → yes (p = 0.97)
Score
Rates on an ordered scale you define: urgency, risk, priority, escalation level.
How urgent is this? → "today" (level 2)
Business use cases
Support triage
Routes every ticket to the right team, detects refund requests and assesses urgency — directly in your CRM, no cloud involved.
Compliance & legal
Classifies sensitive documents and clauses with probability traceability, on data that never leaves your servers.
Business scoring
Prioritizes cases by your criteria (urgency, risk, escalation), recalibrated on your decision history.
Workflow automation
Every decision exposes its probability: define your confidence thresholds and automatic human escalations.
Measured performance
Test accuracy after LoRA fine-tuning (a single epoch), measured on a decontaminated test set. Source: public Unsloth benchmarks.
| Model | Test accuracy | VRAM | Training time |
|---|---|---|---|
| Qwen3.5-0.8B | 78% | 4 GB | 42 min |
| Qwen3.5-2B | 81% | 8 GB | 40 min |
| Llama 3.2 3B | 79% | 4.1 GB | 30 min |
| Gemma 4 E4B | 77% | 14.4 GB | 49 min |
LoRA fine-tuning (r=16, 1 epoch) on 0.8B–4B base models. Training fits on a single GPU.
Our deployment method
Decision audit
Mapping of your recurring decisions and their criteria: what the model must choose, and based on what.
Dataset construction
Training set (state, questions, expected answers) built from your history, annotated and validated with your business experts.
Training & calibration
LoRA fine-tuning of your base model (Qwen, Gemma, Llama), accuracy measured on held-out data and confidence calibration.
Local deployment
Decision API rolled out on your infrastructure — on-premise or air-gapped — and integrated with your business tools.
Why EdgeLLM
Your decision history and training data stay on your infrastructure. The delivered model depends on no third-party API: it runs behind your firewall, with no outbound connection when required.