v3.2
Lumen
Inference, evaluation, and observability, quietly engineered.
The single audited surface for running, measuring, and shipping language model workloads. Built for engineering teams who treat AI as infrastructure, not as a marketing department.
HALDEN-3-70B
412ms
KESTREL-MINI
218ms
ORRIN-2-LARGE
384ms
TALLIS-8X
502ms
VESTA-PRO-1.5
326ms
MERIDIAN-V3
287ms
As shipped — v3.2.1
The console — everything you ship, on one surface
§ 01
The Index
Four primitives. Nothing more.
We don't believe in forty-feature dashboards or AI that does your laundry. Lumen does four things — and does them with the rigor your production workloads quietly deserve.
§ 02
Specimen
A specimen of the type.
Each primitive composed and laid out as in a type specimen — numerals, ascenders, ligatures. The mark of well-set software.
01
Inference, routed.
One SDK across twenty-three providers. Caching, retries, and fallbacks arrive as defaults — not as features wired in over a long weekend.
Provider routing
23 supported
Streaming
SSE · WS
Fallbacks
automatic
Self-hosted models
vLLM · TGI
02
Evaluations, versioned.
Every regression caught before customers find it. Suites build themselves from real traces, prompts and models compared side by side.
LLM-as-judge
built-in
Custom assertions
unlimited
CI integration
GitHub · GitLab
Versioned suites
git-backed
03
Observability, auditable.
Nothing about your system is opaque, including its costs. Filter, replay, export every trace; tokens, latency attach themselves to each request.
Trace retention
90 days
SOC 2 Type II
audited
Self-hosted option
included
Data residency
EU · US · APAC
04
Datasets, annotated.
Every production trace becomes labeled data. Curate, version, and export evaluation sets without leaving the surface that captured them.
Source
production traces
Labeling
inline · API
Export formats
JSONL · CSV
Versioning
git-backed
§ 03
Testimonials
§ 04
Pricing
Honest pricing. No usage traps.
Three tiers, predictable per-seat costs, and a free tier generous enough to ship something real on. The way software pricing used to be, before everything became a metered horror story.