Extract fields from every document — at batch prices.

Invoices, contracts, claims, KYC forms, medical records. Process in the background on European GPUs and get clean JSON back with a per-document audit trail. Run any open-weight vision model, or your own fine-tune.

100K invoices·24h window·pricing at launch
We'll tag your request with structured extraction and pair you with a design-partner slot.
The workload

You have a stream or backlog of semi-structured documents and you need predictable fields out the other side — amounts, dates, parties, line items, clinical codes. Realtime extraction is the wrong shape for this: you don't need answers in 200ms, you need 100,000 documents processed overnight at a sane cost.

Why it's different on sference
Async fit

Backlogs love long windows.

Most extraction jobs are ingest → queue → index. Pick a 24h or 48h window and the orchestrator runs against preemptible spot capacity across EU providers. Document-level checkpointing means a reclaim never costs you more than a single doc's retry.

BYOM

Your fine-tune, same audit trail.

If you've trained a vision model on your customers' contract templates or claim forms, ship it. Anything that runs on vLLM or SGLang we serve — under the same per-document trail as the catalog models.

Compliance

A trail your customer's CISO can verify.

Every document carries model version, infrastructure region, and batch ID. Exportable as JSONL or signed report. No US CLOUD Act exposure.

Example

Example — extract line items from 100,000 invoices.

sference — extraction
$ sference batch ./invoices.jsonl --model qwen3-vl-235b --schema ./invoice.schema.json --window 24h
→ uploading 100,000 records (2.1 GB)
→ batch bch_a12f queued · eta 18h 40m · sla 24h
→ schema validated · 14 fields · line-item array
▸ shard 1/16 · nl-ams · 6,250/6,250 ✓
▸ shard 2/16 · de-fra · 6,250/6,250 ✓
▸ shard 3/16 · fi-hel · preempted → rescheduled
✓ completed 99,984/100,000 · 18h 12m · audit trail exported
input · invoice_0042.pdf
{ "id": "inv_0042", "file": "s3://…/invoice_0042.pdf" }
output · invoice_0042.json
{
  "id": "inv_0042",
  "supplier": "Acme GmbH",
  "vat_id": "DE123456789",
  "total": { "amount": 4812.40, "currency": "EUR" },
  "line_items": [ … 17 items … ],
  "_sference": {
    "model": "qwen3-vl-235b@2026-04",
    "region": "eu-nl-ams",
    "batch": "bch_a12f"
  }
}
SLA & cost

Pricing is announced at launch. The shape below reflects the relative savings — longer windows map to cheaper spot capacity.

pricing at launch
1h

Priority queue, minimal wait.

Baseline
6h

Balanced — typical daytime batch.

Cheaper
24h

The common default for overnight runs.

Much cheaper
48h

Weekend backfills, reprocessing.

Cheapest
Recommended models
Qwen3-VL-235BCatalog

Flagship open VLM; OCR in 32 languages, strong on forms and tables.

Mistral-Small 3.1Catalog

European multimodal option; efficient on cleaner layouts and text-heavy scans.

Your fine-tuneBYOM

Trained on your templates? Same API, same audit trail.

Compliance

Per-document audit trail.

Every extraction carries model version, EU region, and batch ID. Retention is configurable per batch — delete inputs on completion, keep outputs + metadata for your audit log, or retain the lot under a custom policy. Exportable as JSONL or as a signed compliance report.

05Early access

Stop paying realtime prices for work that can wait.

We're in early access. Drop your email — if your workload fits, we'll send you API credentials and you're good to go.