Generate a million samples overnight — with receipts.

Synthetic data is naturally batch. Long windows on spot capacity get you teacher-quality samples at a price that makes regeneration a reflex, not a decision. Every sample carries a model version and batch ID.

1M samples·24h window·pricing at launch

We'll tag your request with synthetic data generation and pair you with a design-partner slot.

The workload

You're bootstrapping a fine-tune, distilling a frontier model into a smaller one, stress-testing a classifier with adversarial prompts, or just filling in a training-set gap. The jobs are embarrassingly parallel and fundamentally non-interactive — yet realtime APIs charge realtime prices and leave you piecing together provenance after the fact.

—Why it's different on sference

Async fit

Generation runs that regenerate cheaply.

Pick a 24h or 48h window and generate at spot-capacity cost. Partial-run checkpointing means a reclaim never loses more than the current shard — the rest resumes automatically.

Teacher choice

Qwen3.6 Plus, Mistral-Medium 3 — or your own teacher.

Open-weight frontier models served on demand, so your teacher only costs you while it's running. BYOM if you've licensed or trained a specific teacher.

Provenance

Model version + batch ID on every sample.

Exportable as JSONL ready for Annex IV documentation. Reproducible regenerations. No silent model swaps between runs.

—Example

Example — generate 1M instruction-tuning samples.

sference — synthetic

$ sference batch ./seeds.jsonl --model qwen3.6-plus --template ./prompt.tpl --window 24h

→ uploading 100,000 seeds (84 MB)

→ batch bch_7c4f queued · eta 19h 10m · sla 24h

→ fanout 10x per seed → 1,000,000 samples

▸ 32 shards · 4 EU providers · streaming output

▸ shard 14/32 · de-fra · preempted → rescheduled

✓ completed 999,412/1,000,000 · 19h 02m · manifest exported

input · seed

{ "id": "seed_0042", "topic": "procurement clauses", "n": 10 }

output · sample

{
  "id": "seed_0042.03",
  "prompt": "Rewrite the following procurement clause…",
  "completion": "…",
  "_sference": {
    "model": "qwen3.6-plus@2026-04",
    "temperature": 0.8,
    "region": "eu-fi-hel",
    "batch": "bch_7c4f"
  }
}

—SLA & cost

Pricing is announced at launch. Long windows are where synthetic generation wins — the cost curve makes regeneration cheap.

pricing at launch

Tight iteration, small sample counts.

Baseline

Same-day data refresh.

Cheaper

24h

Canonical synthetic-data window.

Much cheaper

48h

Full-corpus distillation runs.

Cheapest

—Recommended models

Qwen3.6 PlusCatalog

Flagship MoE teacher; hybrid attention, multilingual, strong reasoning.

Mistral-Medium 3Catalog

European teacher with quality-tier output; solid English + EU languages.

Your teacherBYOM

Licensed or fine-tuned — served under the same audit trail.

—Compliance

Annex IV-ready provenance.

Every sample records teacher model version, sampling params, region, and batch ID. Export the full batch manifest as a signed JSONL and drop it straight into your EU AI Act technical documentation.

—Other workloads

All use cases →

01Workload

Structured extraction

Turn PDFs, scans, and forms into clean JSON — with an audit trail per document.

100K invoices·24h window·pricing at launch

02Workload

Embeddings at scale

Build or rebuild RAG indexes on open-weight embeddings — at a fraction of hosted-API cost.

10M chunks·6h window·pricing at launch

04Workload

Model evals

Run full eval sweeps across base models, fine-tunes, and prompt variants — reproducible and export-ready.

4 models · 10K prompts·24h window·pricing at launch

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.