Backgammon Analysis API: Nardex vs gnubg vs XG

If you are building a backgammon app or analysis tool in 2026, you will choose between three serious engines: Nardex — a hosted neural-net API; gnubg — the long-standing open-source desktop reference; and XG (eXtreme Gammon) — a paid Windows application favoured by tournament players. They solve different problems. This page compares them side by side and tells you which to pick for your use case.

Quick comparison table

When to use Nardex

Pick Nardex if you are shipping a web or mobile product and do not want to bundle an inference runtime. The hosted API returns a full evaluation — equity, win/gammon/backgammon probabilities, ranked alternatives, and a grade label — in a single round trip. There is no ONNX runtime to ship with your client, no GPU to provision, and no model weights to update.

Nardex is also the only option if your product covers long narde (Russian-style narde governed by the Russian Ministry of Sports ruleset). gnubg and XG only handle short backgammon, so you would otherwise have to write your own engine for the long-narde variant.

Use cases where Nardex is the right call: backgammon-as-a-service apps, analysis dashboards, coaching tools, content creators who want automated game review, and research code that needs predictable latency on shared infrastructure.

When to use gnubg

Pick gnubg if your work is academic, requires reproducibility against the published literature, or has to run fully offline under a permissive license. Almost every backgammon paper since 2005 cites gnubg-2-ply rollouts as ground truth, and the source is auditable end-to-end. You can recompile, retrain, and inspect every weight if you need to.

gnubg also wins when you simply do not want to depend on a third-party service. It runs on a laptop without an internet connection, accepts SGF and MAT files, and produces detailed match analysis through a TUI or GTK GUI.

The trade-offs: no API surface, no long-narde support, no modern hardware acceleration, and a C codebase that can be hostile to embed in JVM, .NET, or Node services. Build times on macOS are particularly painful in 2026 because of toolchain drift.

When to use XG

Pick XG (eXtreme Gammon) if you are an individual player, coach, or tournament organiser analysing matches on a Windows machine. The match-equity tables, position editor, and rollout controls are the gold standard for serious cube-action analysis, and many top players publish annotated XG match files.

XG is not a fit if you need automation. The application is closed source, Windows-only, and exposes no public API. There are unofficial scripts that drive the GUI through Windows automation, but they are unsuitable for production services and will break on every release.

Migration guide: gnubg → Nardex

If you currently shell out to gnubg-cli from a server process, the move to Nardex is a single HTTP call. Below is a typical analysis pipeline, before and after.

Before — gnubg subprocess:

# Spawn gnubg, pipe XGID, parse text output
echo "$xgid" | gnubg -t -q --commands "set evaluation plies 2; analyse position"

After — Nardex API:

curl -X POST https://nardex.ai/api/v1/analyze \
  -H "Authorization: Bearer $NARDEX_KEY" \
  -H "Content-Type: application/json" \
  -d '{"game_type":"backgammon","position":{...},"decision":{...}}'

The response is structured JSON: equity, win/gammon/backgammon probabilities, ranked alternatives, and a grade. No text parsing, no subprocess management, no GIL-blocking calls. See the Nardex Analysis API docs for the full request and response schema.

Benchmarks and validation

A full latency and accuracy report — including ply-zero p50/p95 timings and equity-error distribution against gnubg-2-ply baseline — is planned for publication alongside the public API release. Until then, expect hosted-API call times to be dominated by network round-trip from your region rather than inference itself; CUDA inference of a single position is sub-millisecond on contemporary GPUs.

For backgammon specifically, the Nardex network architecture follows the lineage of TD-Gammon and gnubg's neural network: positional input encoding, multi-layer feed-forward, six output probabilities (win, gammon, backgammon for each side). This design has decades of empirical validation as a backgammon evaluator.

Long-narde validation is more challenging because no external reference engine exists. Internally, Nardex networks were validated against hand-curated expert positions from the Russian Ministry of Sports rulebook plus self-play tournament games per training generation. See backgammon basics for context on how rollout-based training works.

Frequently asked questions

Is gnubg still maintained?

gnubg receives sporadic patches but its last full release was in 2023. Development is volunteer-driven, the codebase is C from the mid-2000s, and there is no public roadmap. It still works for offline analysis on Linux and macOS, but expect quirks with modern build toolchains.

Does XG (eXtreme Gammon) have an API?

No. XG is a closed-source Windows desktop application. It exposes match-file import/export and a proprietary scripting interface, but there is no documented HTTP API or SDK. To use XG programmatically you have to drive the GUI with automation tools, which is not viable for production services.

What is Nardex's accuracy compared to gnubg?

Independent benchmarks against gnubg-2-ply are forthcoming and will be published after the public API release. Internally, Nardex networks were validated against expert-played positions and self-play tournament games. On long narde — which gnubg does not support — Nardex is the only modern engine with neural-network evaluation.

Can I self-host Nardex?

Self-hosting is on the roadmap but not yet generally available. The hosted API runs in the EU. If you need on-prem deployment for regulatory reasons, mention it in your access request.

How is Nardex priced?

Nardex is currently in private beta. Pricing tiers will be announced ahead of public release. To request beta access, leave your contact details on the /developers page.

Which engine is best for academic research?

gnubg remains the standard reference for published backgammon research because of its open-source TD-Gammon-style network and reproducible rollouts. Use gnubg as a baseline; use Nardex when you need an HTTP API, support for long narde, or modern hardware acceleration (CUDA, TensorRT).