GPU Util — % utilisation
GPU Temp — °C die
Unified — GB of 128 · 8 GB guard
Throughput — tok / second
TTFT — ms · first token
throughput & first-token from the active lane
Active Lane idle no warm brain

← Models

arena_run · astro

An operator cockpit you run on your own DGX Spark

Public model arenas optimise for crowdsourced vote volume on cloud APIs and never know what hardware their votes ran on. Orionfold Arena is the operator-driven, local-first alternative: a one-process cockpit that drives every quant, LoRA, bench, harness, and skill your Spark has produced — private eval leaderboards, a cost-per-quality efficiency frontier, side-by-side rubric-scored compare, live telemetry, and a co-iteration Lab. Your prompts, telemetry, and eval data never leave the box.

base fieldkit[arena] · Astro + FastAPI sidecar · license apache-2.0

▶ Try in chat ＋ Send to compare
What it's for
  • Chat against the resident brain with live tok/s + unified-memory telemetry
  • Side-by-side rubric-scored compare (resident lane vs OpenRouter frontier)
  • Private leaderboards + a cost/quality Pareto frontier anchored to real Spark hardware
  • Browse the model/notebook/bench/harness catalog with quant economics
  • A Lab board for the operator+AI build loop, with a private notes margin

Audience — DGX Spark operators (GB10, 128 GB unified memory) and edge-AI builders who want a beautiful, local-first cockpit over their own artifacts — no hosted SaaS, no data egress.

Known drift bounded · honest
  • Live surfaces (chat, compare, telemetry, Lab notes) need the local sidecar On the public web preview the sidecar is offline by design; ~5 surfaces degrade to their static mirror data + an offline banner. The leaderboard, models browser, frontier, and Lab board render fully static.