Pipelines that are reproducible by default
Declarative stages, pinned dependencies, and content-addressed artifacts mean a run from six months ago reproduces bit-for-bit today.
Versioned configs
Time-travel runs
Build the next generation of AI systems with Seufic.
A complete, local-first stack for AI pipelines, dataset creation, fine-tuning, evaluation, and inference — engineered to give developers full control, with no vendor lock-in.
Open-source first
Local inference
Your data stays yours
Products
Everything you need to ship AI
A coherent toolchain that replaces a dozen disconnected services — from raw data to deployed models.
AI Training Pipeline
Reproducible training runs with versioned configs, checkpoints, and metrics — orchestrated from one place.
Dataset Builder
Curate, clean, label, and version datasets with deterministic transforms and full lineage tracking.
Model Fine-Tuning
Adapt open models with LoRA, full fine-tuning, and parameter-efficient methods out of the box.
Evaluation System
Benchmark quality, regressions, and safety with custom suites that run on every model revision.
Local Inference Engine
Serve on your own hardware with low-latency batching, quantization, and an OpenAI-compatible API.
Open-Source First Stack
Every layer is inspectable and replaceable. No black boxes, no lock-in — extend it however you need.
Datasets you can trust and trace
Every transform is recorded, every split is deterministic, and every row carries its lineage — so you always know exactly what your model learned from.
Full lineage
Deterministic splits
Inference that runs where you do
An OpenAI-compatible local engine with batching and quantization. Ship to a laptop, a workstation, or a cluster — same API, your hardware.
Low latency
Self-hosted
$seufic serve --model llama3-8b --port 8080
→ Loading model weights…
✓ Quantization: 4-bit (Q4_K_M)
✓ Engine: continuous batching enabled
✓ API listening on 0.0.0.0:8080
──────────────────────────────
$curl localhost:8080/v1/chat/completions \
-d '{"model":"llama3-8b","messages":[…]}'
↳ 38ms · 312 tok/s · batch 4
$
About us
Built for engineers who want control
Seufic exists for teams who believe AI infrastructure should respect ownership, transparency, and portability — from the first experiment to production scale.
Open-source first
The whole stack is open and auditable, so you can trust, fork, and contribute to what you depend on.
Local-first development
Run end-to-end on your own machines. No mandatory cloud, no data leaving your environment.
No vendor lock-in
Standard formats and open APIs mean you can move on any time. Your infrastructure is never a hostage.
Developer-focused architecture
A clean CLI, typed SDKs, and composable primitives that fit existing engineering workflows.
Updates
What's new at Seufic
Releases, research notes, and product changes — shipped continuously.
Jun 18, 2026
Release
Local Inference Engine 2.0
Up to 2.4× higher throughput with continuous batching and 4-bit quantization on consumer GPUs.
Jun 02, 2026
Product
Dataset lineage graph
Visualize every transform from raw source to training split, with one-click reproduction.
May 21, 2026
Open source
Evaluation suites, open-sourced
A library of reproducible benchmarks for quality, safety, and regression testing.
May 09, 2026
SDK
Typed SDKs for Python & TypeScript
First-class clients with full type coverage across the pipeline, dataset, and inference APIs.
Yes. Every layer of the stack — pipelines, dataset tooling, evaluation, and the inference engine — is open and auditable. You can self-host, fork, and contribute back.
No. Seufic is local-first. You can run the entire workflow on your own machines, and your data never has to leave your environment.
Any open-weight model. We support LoRA, full fine-tuning, and parameter-efficient methods, with checkpoints versioned alongside your datasets.
The local engine exposes an OpenAI-compatible API, so existing clients work with a single base-URL change — while everything runs on hardware you control.