Fabryka · Operations · Vol. IAggregate production telemetryRolling 30 days · UTC
§ 00 · Live production evidence
How the factory
is working.
Completed requests and processed tokens, aggregated by week and model. Prompts, responses, API keys, and customer identities are never included.
Today · UTC · in progress
Today so far.
Completed tasks since 00:00 UTC.
Coverage: recorded token totals for completed tasks. Historical batch metrics exclude synthetic accounts; the Basal meter includes API, public demo and verification calls. In-flight and failed tasks are not included here. Basal includes completed API and anonymous demo calls from its new aggregate meter; input tokens are counted and output tokens are zero. Earlier Basal requests recorded as batches are retained. Cached-input and reasoning-token breakdowns are not currently retained, so cache savings cannot be shown separately.
Weekly traction
Seven complete days.
Excludes today · loading summary…
Capacity utilisation
Usage as a % of capacity.
Compare the last seven complete UTC days with a ten-minute capacity measurement, extended to continuous operation for seven days.
Measured on the production Qwen line for 10 minutes: 39,816 input + output tokens, 316/316 successful requests. Seven-day capacity = measured ten-minute output × 1,008. This is workload-dependent token capacity, assuming 24/7 availability, not GPU utilisation.
0%100% capacity
No maximum-capacity benchmark has been supplied.
Simple trend extension
If this daily pace holds.
Current-week tokens divided by observed days, extended to seven days.
Rolling 30-day totals
§ 01 · Weekly production
Tokens processed
Monday–Sunday · UTC. Weeks with recorded usage below $0.01 are hidden, except weeks with Basal activity. Partial weeks are marked. Dashed extensions are simple seven-day linear projections; solid bars are actuals.
Bielik 11B v3Qwen3.8 27BPLLuM 12BQwen + BielikBasalMixed batches
>Linear projection
——
§ 02 · Production lines
Model-level operations.
Throughput is aggregate output tokens divided by recorded execution time. It is workload-dependent and should not be read as a controlled benchmark.
| Model | Requests | Tokens | Average response | Output tok/s | Usage |
|---|
§ 03 · Methodology
What this report proves.
Only completed Bielik, Qwen, PLLuM, hybrid, and mixed requests are counted. Token counts and durations come from the serving metadata written when a request completes.
The public endpoint is refreshed every minute and returns a rolling 30-day UTC series, summed here into Monday–Sunday weeks. The weekly summary compares the last seven complete UTC days with the preceding seven; today is excluded from that comparison. The first and current chart weeks may be partial. The report intentionally excludes raw prompts, generated content, credentials, IP addresses, and account-level records.