Five top-level views, plus three you reach from them. The header always tells you where you are, which box you are judging against, and how fresh the data is:
INFERENCE ATLAS [target] · box NVIDIA DGX Spark (GB10) · data @ 46e00af · fresh · checked 6m ago
▲ ▲ ▲ ▲ ▲
view target box (b to change) data commit sync age
fresh means the last check found nothing new. updated means shards were re-fetched.
offline (cache) means the check failed and you are reading the cache — everything still
works.
1 · Target
The launch screen, and the one that answers "what should I run here?".
Your box — what the app knows about the target: the registry id, memory, bandwidth, and whether it was detected on this machine or chosen by you.
Worth running on \<box> — every measurement in the atlas, ranked. Fit first
(✓ recommended before ~ should fit before ! tight before ✗), then by the site's
headline metric within each group. Columns:
| Column | |
|---|---|
| fit | the verdict for your box — see The target box |
| model | Hugging Face repo id of the model |
| quant | the quantization id within that model |
| engine | engine and exact version |
| workload | the pinned workload definition that was run |
| headline | the metric that matters for that kind of run |
The footer counts what the atlas holds: runs, models, hardware, engines.
2 · Runs
The same measurements, unranked and unfiltered — the whole table, with the identity columns
in full and / to narrow it.
Use this when you know what you are looking for: a model, an engine, a piece of hardware you do not own. Unlike the target view it does not sort by fit, so it is also how you see what other people are running.
3 · Pareto
Throughput against latency for every serving run: x is TTFT p50 (lower is better, on a log scale because the range spans milliseconds to minutes), y is output tok/s (higher is better).
The frontier — the set of runs that nothing else beats on both axes at once — is drawn
bright and connected. Everything else is a dim dot. Walking with j/k moves the ◉
marker and names the run underneath.
This is the view for "what am I giving up?". A configuration far off the frontier is being beaten by another one on both axes, and the panel underneath tells you which.
4 · Coverage
Model × hardware, one coloured cell per square, brightness by how many runs exist. Blank means nobody has measured it — which in this atlas is the interesting part, not a gap in the UI.
The legend under the map numbers the hardware columns, since device ids are too long to sit above three-character cells.
This is the contribution view: a blank square is twenty minutes of work that nobody has done
yet, and AGENTS.md explains how to fill one.
5 · Hardware (also b)
Pick the box every verdict is judged against, and how many devices you have. Fully covered in The target box.
Run detail (enter)
Everything known about one measurement.
Identity — model, quantization, engine and version, the hardware it ran on and how many
devices, the workload, who ran it, and the verification level (self-reported until someone
else reproduces it). Underneath: only the headline metrics that run actually carries, so an
eval shows accuracy and an evaluation of a serving run shows tok/s and TTFT.
Fit on \<your box> — the verdict, framed in its own colour, with every reason listed. This is the reasoning, not a score: what was measured, what was estimated, what it means for your hardware.
Latency distribution — TTFT p50 / p90 / p95 / p99 as bars from zero, with the colour deepening towards p99 so a tight tail is still readable.
Requests — every request in the run as one column, tallest = slowest, resampled across the whole run so nothing is hidden. Failures are red. This is where you see a run warm up, or a queue building, or one call that hung.
Sweep — for sweep and long-context runs, throughput (filled, cobalt, left axis) against latency (orange, right axis) over the swept variable. The x axis is labelled with the levels that were actually run.
Gotchas — one framed note per gotcha, coloured by severity: red for a blocker, orange for a warning, grey for a note. This is the institutional knowledge that outlives the number — a flag whose default is a lie, a container tag that only exists for one architecture, a parser name that resolves under one spelling.
Press g here to turn all of it into a recipe.
Recipe (g from the detail)
The generated Markdown, scrollable, with the file path it was written to at the top and your configured agents listed underneath. Covered in Recipes.
Help (?)
The keyboard reference. ? again, or esc, returns you to where you were.