Your experiments, in one file. A single Go binary. MLflow-API-compatible. First-class LLM trace support. Zero databases. Zero config. litemlflow up and you're tracking.

• 143× faster cold start than MLflow — 53 ms vs 7.5 s. No Python interpreter, no ORM, no migration dance in CI.
• One env variable to migrate — MLFLOW_TRACKING_URI=http://localhost:5000. Your existing MLflow code keeps running.
• LLM-native from day one — traces, prompts (versioned + aliased), evals, and OpenTelemetry OTLP live beside classic ML metrics in the same UI.

# Raw binary (Linux / macOS amd64+arm64) — recommended
curl -fsSL -o /usr/local/bin/litemlflow \
  https://github.com/gorevds/litemlflow/releases/latest/download/litemlflow-linux-amd64
chmod +x /usr/local/bin/litemlflow
litemlflow up --data ~/lmf

# Docker
docker run -p 5000:5000 -v $(pwd)/data:/data ghcr.io/gorevds/litemlflow:latest up

# Kubernetes (Helm)
helm install lmf oci://ghcr.io/litemlflow/charts/litemlflow --version 0.1.0

# Build from source (Go 1.26+)
git clone https://github.com/gorevds/litemlflow && cd litemlflow && make build

Snap, Homebrew, Debian, RPM, and the Terraform provider were sunset in v1.2. See dist/_sunset/README.md for the rationale and supported channels.

litemlflow up --data ./data          # server on :5000 in ~53 ms

import mlflow
mlflow.set_tracking_uri("http://localhost:5000")
mlflow.set_experiment("my-project")

with mlflow.start_run():
    mlflow.log_param("lr", 0.01)
    mlflow.log_metric("loss", 0.42)

Open http://localhost:5000 — you're done. No Postgres. No S3. No reverse proxy.

Benchmark details: docs/bench-v04.md. All numbers measured on the same loopback machine.

LiteMLflow follows semver starting at v1.0. v2.0 is the LTS stream.

• The HTTP wire shape of /api/v1/... and /api/v2/... is frozen for v2.x. Additive changes only (new optional fields/endpoints) — see ADR 0003.
• v1-flagged legacy MLflow-compat aliases ship with Sunset: Tue, 11 May 2027 00:00:00 GMT headers — 12 months of advance notice.
• Pin to /api/v2/... paths in new code if you want the explicit LTS namespace stamp; existing v1 clients keep working unchanged.

Federation caveat (/api/v1/federate/...): in-tree independent reviews caught and fixed RBAC bypass (C1), unbounded-response OOM (C2), peer-name enumeration (H2), and cache-correctness (H1) issues. An external pen test on the federation protocol has NOT been performed. Deploy federation only across servers you control (single trust domain). Cross-organization federation should wait for the pen-test pass tracked in docs/upgrade-to-v2.md.

• MLflow REST API — experiments, runs, metrics, params, tags, artifacts (upload/download/delete), metric history, search with full filter grammar (=/!=/</>/LIKE/IN/BETWEEN), log_batch, log_inputs, pagination, Model Registry
• Native API — traces (manual + OTLP HTTP/gRPC), prompts (versioned, content-addressed, aliased), evals
• Embedded UI — experiments → runs → run detail; metrics charts, trace waterfall, prompt diff, bulk compare/export/tag; run notes (markdown), starring, inline rename, artifact preview; side-by-side run comparison (params diff, metrics overlay, tags, run summary); custom column picker (per-experiment, persisted); share button (copies current URL); keyboard shortcuts; global search (⌘K palette finds runs/experiments/prompts cross-experiment); dark/light theme; embed mode
• Auth — anonymous, basic, OIDC (PKCE + RS256 + JWKS), session cookies, RBAC (viewer/editor/admin)
• Storage backends — filesystem (default, zero config) and S3-compatible (MinIO/AWS, pure-Go SigV4, multipart)
• Observability — Prometheus /metrics endpoint (12 metric families), server-side LTTB downsampling
• CLI — up, migrate, rollback, backup, restore, import-mlflow, version
• Distribution — Docker image, Helm chart, Kubernetes operator, raw binary (Snap / Homebrew / Debian / RPM / Terraform sunset in v1.2 — see dist/_sunset/)
• Python SDK — litemlflow[langchain] (auto-instrument LangChain), litemlflow[llamaindex] (auto-instrument LlamaIndex)
• Tested against real MLflow Python client 3.x — 31/31 compat checks pass live at https://lmf.gorev.space

# Before
mlflow.set_tracking_uri("http://mlflow-server:5000")

# After
mlflow.set_tracking_uri("http://litemlflow-server:5000")
# Everything else stays the same.

To import existing data from a running MLflow instance:

litemlflow import-mlflow \
  --src http://mlflow-server:5000 \
  --data ./data

cmd/litemlflow/      CLI entry point
internal/
  api/mlflow/        MLflow REST compat layer
  api/native/        LiteMLflow native API (traces, prompts, evals, OTLP)
  artifact/          Filesystem + S3 artifact storage
  store/             SQLite store (WAL mode, LTTB downsampling)
  server/            HTTP/2 server, middleware, auth, RBAC, metrics
  migrations/        Embedded SQL migrations + runner
python/litemlflow/   Python SDK + LangChain + LlamaIndex handlers
operator/            Kubernetes operator (separate Go module)
dist/                Helm chart (active); _sunset/ holds Snap/Brew/RPM/Deb/Terraform (unmaintained, see README inside)
docs/                Human-facing documentation
tests/integration/   End-to-end tests (real MLflow client + bench harness)

make build          # → bin/litemlflow (requires Go 1.22+)
make test           # Go + Python tests
make compat-test    # real MLflow Python client against local binary
make lint           # static analysis
make fuzz-short     # fuzz targets (parsers, JWT, OTLP)

LiteMLflow is free and Apache 2.0. If it saves you time, consider sponsoring development:

• GitHub Sponsors (TODO: activate before launch)
• Polar.sh (TODO: activate before launch)

See CONTRIBUTING.md. DCO sign-off required (git commit -s); no CLA — your copyright stays yours.

Apache 2.0 — see LICENSE and NOTICE.

LiteMLflow is independent of and not affiliated with Databricks, Inc. References to "MLflow" describe API compatibility only.