Small models. Big operational savings.
Sub-10B models tuned for one job, running on a single GPU, a CPU or an edge device, with semantic guardrails and a verification harness. For workloads where latency, cost, privacy or resource limits rule out a frontier model.
Four reasons to go small.
Teach it the job, not the internet.
A small model cannot know everything, so it should know one thing very well and refuse the rest cleanly.
Distil from larger models
Use larger models to generate and verify training data for the task, then train the small model on the verified set.
Verify against ground truth
A verification harness tests the small model against ground-truth files with confidence scoring, so you know where it is safe.
Guardrails and structure
Semantic guardrails, structured outputs and calibrated refusal for anything outside the task.
Serve where it fits
vLLM on a GPU, llama.cpp on a CPU, MLX on Apple silicon, or an edge box. Monitored on the cockpit like everything else.
A 7B that holds a brand voice on one GPU.
B2B Marketing (7B)
The compact first release. The same B2B copy discipline as the 32B flagship, light enough for a single modest GPU.
- LoRA adapter on Qwen2.5-7B-Instruct · Apache-2.0
- 13,000+ curated, brand-scrubbed training pairs
- Zero-leak verified across generation tests
- Suited to on-device and single-GPU serving
Where small wins
- Classification, routing and extraction at high volume
- NL2SQL and structured queries over your own schema
- On-device coaching and drills, such as the audio replay in KshetraIQ
- Guardrail and policy checks in front of a larger model
- Anything that must run offline, in a plant, a branch or a vehicle
Common questions.
How small is small?
Usually 0.8B to 8B parameters. The size follows the task and the hardware it must run on, not the other way round.
Will it be as good as a frontier model?
On the one job it was tuned for, often better, because it has seen your data. Outside that job it should refuse, and we test that it does.
Can it run fully offline?
Yes. That is one of the main reasons to choose one. We package the model, the runtime and the guardrails together.
Shrink the model, keep the quality.
Bring the workload and the hardware you have. We will tell you whether a small model can carry it. No pitch deck. Just a technical conversation.