Experiment tracking that keeps up with training.

Log runs, compare metrics, inspect artifacts, and export your data without waiting on the dashboard.

Python SDKW&B, MLflow, Neptune importsHosted + Premium BYOC ClickHousePredictable pricing
Runs100 selected
hosted-scale-cpr_8f21 · 188.95
hosted-scale-cir_77ac · 198.82
sweep-baseliner_3c18 · 186.76
eval-regressionr_52de · 199.49
Line chart · eval/return_meanReturn Mean
GroupedFull fidelityStep 124
020k140200
Large run history
50k runs
Newest-page reads are 236 ms p95.
Best-run sorting
<1 sec
Sort by metric best from maintained summaries.
Fast chart opens
<1 sec
1,000 chart points read in 224 ms p95.

Install the SDK. Log the run.

The Python SDK buffers metrics and sends them to a Rust + ClickHouse backend. Run lists stay summary-backed. Charts fetch bounded series.

01

Log from your loop

The Python SDK buffers init, log, and finish straight from the trainer, and spools offline when the network drops.

PyTorchJAXTRL
02

Rust + ClickHouse hot path

Metrics land as typed summaries with indexed search, so run lists and best-metric sorting stay sub-second on large projects.

typed summariesindexed search
03

Read, compare, export

Open charts, compare runs side by side, and pull everything back out. Artifacts and checkpoints live in your own S3 / R2.

chartscompareS3 / R2

The parts teams use every day.

Metric charts

Readable curves for selected runs.

Streamed scalar series, grouped averages, smoothing, zoom, and hover readouts.

run · llm-7b-sft · loss
streaming
train
eval
Benchmarked reads

Fast reads on large projects.

Current hosted benchmarks keep a 50k-run project sub-second for newest pages, best-metric sorting, and 1,000-point chart reads.

Newest runs page p95
236
ms · hosted 50k runs
SDK queue

Buffered logging.

init · log · artifact · checkpoint · finish.

sdk.tail · run r_a4e2
tailing
12:42:58init cfg=24
12:39:47loss=1.82 lr=2e-4
12:36:36artifact png 128KiB
12:33:25loss=1.74 grad=.94
12:30:14ckpt model.pt 12.4GiB
12:27:03flush 2048/2048
12:24:52loss=1.66 vram=78GiB
12:21:41finish 4h12m 18k ev
12:18:30init cfg=24
12:15:19loss=1.82 lr=2e-4
12:12:08artifact png 128KiB
12:09:57loss=1.74 grad=.94
12:06:46ckpt model.pt 12.4GiB
12:03:35flush 2048/2048
12:00:24loss=1.66 vram=78GiB
12:57:13finish 4h12m 18k ev
Run data

Typed fields.

Configs, metrics, tags, artifacts, histograms, and media stay queryable.

floatsstringsfileshistogramstags
Migration

Bring old runs with you.

Import W&B, MLflow, and Neptune JSON. Shadow W&B scalar logs during migration.

W&B JSONMLflow JSONNeptune JSONshadow_wandb

Three calls cover the loop.

The SDK is small on purpose: init, log, and finish cover the daily loop. Artifacts and checkpoints are files.

Python 3.11+Rust APIClickHouseOpen SDK
train.py
$ pip install instantml && python train.py
# Start a run, log metrics, finish.
import instantml as im

run = im.init(project="llm-7b-sft", config=cfg)

for step, batch in enumerate(loader):
    loss = train_step(batch)
    run.log({"loss": loss}, step=step)

run.log_checkpoint_file("./ckpt/model.pt", step=step)
run.finish()

Bring one project.

We'll help you import it, compare real runs, and decide if InstantML earns a place in your training loop.

hello@instantml.ai