Quasar

Training active

>_Quasar Training

Almost one month into this run, on relatively low compute, Quasar is already showing measurable benchmark movement. The training is live, verified, and ready to scale.

Quasar Preview Subnet 24 Elastic Training
01

Current Run

Confirmed run status, elapsed time, and benchmark movement.

Run Age≈1 monthCurrent production training run
Tokens Processed1.3TEstimated during the current run
Benchmarks Up7 / 9Largest movement: ARC Easy +3.86
02

Benchmark Movement

Old Quasar compared with the current model after roughly ten percent of this training run.

7 / 9Benchmarks moving up

Almost one month into the run, with relatively low compute, Quasar is already showing broad benchmark movement. ARC Easy leads at +3.86 points. As compute support grows, the training can scale further.

MMLU

+0.61
68.40→69.01
Old → current

MMLU-Pro

+0.52
33.20→33.72
Old → current

GPQA

+0.41
25.60→26.01
Old → current

ARC Challenge

+1.10
63.00→64.10
Old → current

ARC Easy

+3.86
80.10→83.96
Old → current

PIQA

−0.40
81.90→81.50
Old → current

HellaSwag

+0.79
74.00→74.79
Old → current

OpenBookQA

−0.60
47.00→46.40
Old → current

MATH-500

+0.86
71.40→72.26
Old → current
03

Training Loss

Measured learner loss and released checkpoint evaluations from the current run.

Verified training telemetry

Training Loss

Awaiting verified telemetry n/a

Awaiting verified loss telemetry

PolicyNo synthetic curve Raw samples0 Displayed points0 ProcessingUnmodified
04

Run Context

Only confirmed information from the current training update.

TrainingActive
ElapsedAlmost one month
Compute profileRelatively low
Scale stateReady to expand