live eventEidos is moving itself to 90%-cheaper silicon without losing intelligence. Watch.follow on LinkedIn →
local authorship69%$0.48savedhow?
Mission: local silicon authors 90 percent of event narration. Currently 69 percent. Hosted cost incurred $0.21. Savings $0.48.

run detail

thunder-h100-3800c438c8

H100 eval suite (backfill from profile.json)

started 4h agoended 1h agoduration 153mstatus completedsession thunder-backfill

lanes

gputok/slatency (ms)$/hr$/1M tokens
H100winner154.7742$2.49$4.47

eval scores · 120

Rubric-scored quality measurements on this run's model outputs. Higher composite = better.

modeluse casetestcompositetok/s
qwen2.5:14bchunkingchunk_technical_doc94.02.6
qwen2.5:14bchunkingchunk_technical_doc94.03.1
qwen2.5:14bchunkingchunk_mixed_content92.02.7
qwen2.5:14bchunkingchunk_mixed_content92.02.9
qwen2.5:14bchunkingchunk_code_narrative94.03.0
qwen2.5:14bchunkingchunk_code_narrative94.02.8
qwen2.5:14bchunkingchunk_short_text97.03.5
qwen2.5:14bchunkingchunk_short_text97.03.4
qwen2.5:14bsearch_querysq_temporal_filter84.03.4
qwen2.5:14bsearch_querysq_temporal_filter84.03.4
qwen2.5:14bsearch_querysq_code_search100.03.4
qwen2.5:14bsearch_querysq_code_search84.03.0
qwen2.5:14bsearch_querysq_multi_source96.32.9
qwen2.5:14bsearch_querysq_multi_source96.33.0
qwen2.5:14bsearch_querysq_memory_recall100.03.1
qwen2.5:14bsearch_querysq_memory_recall100.03.0
qwen2.5:14bsearch_querysq_delta_search84.03.3
qwen2.5:14bsearch_querysq_delta_search84.03.2
qwen2.5:14bcontext_synthesissynth_architecture93.23.2
qwen2.5:14bcontext_synthesissynth_architecture92.63.1
qwen2.5:14bcontext_synthesissynth_dietary87.03.5
qwen2.5:14bcontext_synthesissynth_dietary87.03.6
qwen2.5:14bcontext_synthesissynth_conflicting67.03.5
qwen2.5:14bcontext_synthesissynth_conflicting87.03.5
qwen2.5:14bmemory_extractionmem_dietary90.93.2
qwen2.5:14bmemory_extractionmem_dietary94.73.3
qwen2.5:14bmemory_extractionmem_incident100.03.3
qwen2.5:14bmemory_extractionmem_incident100.03.4
qwen2.5:14bmemory_extractionmem_preferences100.03.3
llama3.1:8badapter_extractionadapt_email100.04.7
llama3.1:8badapter_extractionadapt_email100.05.4
llama3.1:8badapter_extractionadapt_imessage100.05.4
llama3.1:8badapter_extractionadapt_imessage100.05.8
llama3.1:8badapter_extractionadapt_code_file58.05.2
llama3.1:8badapter_extractionadapt_code_file58.06.4
llama3.1:8badapter_extractionadapt_voice_memo100.06.2
llama3.1:8badapter_extractionadapt_voice_memo100.06.4
llama3.1:8bclassificationcls_email100.06.4
llama3.1:8bclassificationcls_email100.06.8
llama3.1:8bclassificationcls_imessage100.06.5
showing 40 of 120

boothH100 just posted 591 tok/s on llama3.2:1b — 24.2× faster than the A6000 lane (2m ago).

Crucible — live.eidosagi.com