Thermitev0.1.3

Case studies

Campaigns Thermite has run, from the first question to the last verdict, with the judge's own numbers.

Thermite tensorless

Qwen3.8-27B on a $100 GPU, 3.9× faster

Four campaigns tuned an inference engine for Qwen3.8-27B on one Tesla P100: two goal campaigns made the new weight format work and hold its quality, then two hill climbs made it fast, with the model's answers held where they started. tensorless ships the result in one command.

3.89×
faster decode, 7.74 → 30.06 tokens a second
131
changes judged in the decode climb, 28 kept
$100
the GPU: one used Tesla P100
Read the studyRun it with tensorless
Decode speed of every attempt, in tokens a secondQwen3.8-27B-PN3 on one Tesla P100, 24–29 October 2026
A replay of Thermite's decode hill climb on Qwen3.8-27B-PN3, on one Tesla P100: 131 changes judged in five days. 28 were confirmed wins, 48 were slower, 39 made no real difference, 12 broke a guard and 4 were inconclusive. The champion's decode speed rose from 7.7 to 30.1 tokens a second, past Q3Floor's 26.3.

All case studies

AI and machine learning

3.89×

faster decode on one Tesla P100

Qwen3.8-27B on a $100 GPU, 3.9× faster

Four campaigns tuned an inference engine for Qwen3.8-27B on a $100 GPU: 131 changes judged in the decode climb, 28 kept, and the model's answers held where they started.

goalhill climb

Local AI · Startups · with tensorless

Read the study

Share a campaign

Ran a campaign worth a case study? Bring its results to the Hackers in the Loop Discord.

Join the DiscordSee the use cases