Your weight.Your own endpoint.

Push fine-tunes, LoRAs or fully custom checkpoints. We serve them behind a private endpoint nobody else can address - at serverless prices.

Private endpoints · never shared . never pooled

Model registryPrivate

support-triage-v4ready

Qwen 3.8 27B · LoRA · fp8

claims-extractorready

Gemma 4 31B · full fine-tune

code-reviewer-2building

DeepSeek V4.1 Flash · fp8

Reachable only from your organisation

Adapter support: LoRA
Cold start: 60 s
Endpoint access: private

What you get

A private model, without a private cluster.

No Dockerfile · no CUDA matrix

Containerised on upload

Push a checkpoint or a LoRA and we build a sized, quantised image for the GPU it will run on. You skip the serving stack entirely.

Private by construction

Every custom model gets its own endpoint inside your org. No other tenant can address it and no prompt routes to a shared pool.

Same API, your weights

Your fine-tune answers on the same OpenAI-compatible path as the public catalog, with per-model rate limits and metrics.

Upload

Bring weights. Get an endpoint.

One command takes a checkpoint from your laptop to a served endpoint. We scan it, build the runtime image and wire it into the same autoscaler as every other model.

LoRA adapters merge into the base at build time
fp8 or int4 quantisation chosen for the target GPU
Private endpoints autoscale exactly like catalog models

abhigya models push

private

abhigya models push ./lora-out \

--name support-triage-v4 \

--base qwen/qwen3.8-27b \

--private

Scan passed · image built in 1m 42s

Quantised to fp8 · endpoint ready

Endpointhttps://api.abhigya.ai/v1/models/support-triage-v4