ON1 . AEDL benchmark
landing . fitness . ops . docs

AEDL sign-off

topic
[approval-aedl-model-benchmark-v0]
approve baseline deny / amend Akash A100 40GB baseline: 7 models, 93.2 GB estimated download

gate before pulls

defaultbaseline
statusapproved; waiting remote host
costAkash/Vast GPU; Njalla/FlokiNET control baseline
executeblocked until AEDL_REMOTE_HOST

AEDL rule offspring

targetstability, capability, profitability
split30% parent default, optimized per aegent
spawnonly if 30-day lower bound is profitable
migratetrigger at 25% backend cost improvement or SLA failure

baseline pull list commercial-use screened

rank model tag license size role
1gpt-oss 20Bgpt-oss:20bApache-2.0 + policy14.0 GBgeneral reasoning and tool-use baseline
2Qwen3-Coder 30B-A3Bqwen3-coder:30bApache-2.019.0 GBrepo-scale coding and implementation
3Qwen3 30B-A3Bqwen3:30bApache-2.019.0 GBefficient multilingual generalist
4Mistral Small 3.2 24Bmistral-small3.2:24bApache-2.015.0 GBEuropean long-context generalist
5DeepSeek R1 Distill Qwen 32Bdeepseek-r1:32bMIT/distill notice20.0 GBdeliberate reasoning and critique
6Phi-4 Mini Reasoningphi4-mini-reasoning:latestMIT3.2 GBsmall-reasoner control
7SmolLM3 3Balibayram/smollm3:latestApache-2.03.0 GBtiny open control

commands after approval

python3 scripts/send-aedl-model-questionnaire-email.py
python3 scripts/mail-verdict-poller.py --since-hours 24
python3 scripts/aedl-remote-executor.py --selection baseline --backend akash --sku a100-40gb --json
export AEDL_REMOTE_HOST=user@fresh-gpu-host
python3 scripts/aedl-remote-executor.py --selection baseline --backend akash --sku a100-40gb --execute