wave inference · model routing plane

WAVE leaderboard

what we measured·what actually ran

—completions served
—metered spend
—tokens in + out
7models on the plane
01 /

measured quality

The served plane, scored against a frozen eval snapshot. Quality and pass-rate are separate metrics and are not interchangeable. Anything unmeasured is marked unmeasured — never filled in with a guess.

Model Backend Metric Quality Taskset
qwen2.5:3b edge judge_consensus
0.8
✓ 5/5
qwen3-coder:30b runpod judge_consensus
1
✓ --
gpt-oss:120b runpod —
unmeasured
— --
qwen3-coder:30b-pool local taskset-pass-rate
1
✓ 5/5
gemma4:31b local taskset-pass-rate
0.8
✓ 4/5
wave-qwen38-ad-iq3s local taskset-pass-rate
0.8
✓ 4/5
deepcoder:14b local taskset-pass-rate
1
✓ 5/5

snapshot 2026-08-20 · baked eval, not live
judge_consensus agreement across a judge panel · taskset-pass-rate deterministic pass count

local tier · live on the rigs

models serving on WAVE's own hardware — measured, receipted, $0/token

modelbackendlatencycapabilitiesscore
glm-5.3-744b-iq2m llama.cpp-b10665-glm-dsa batch batch-eval, synthetic-data, judge 21/21 wave-core
laguna-xs-2.1-apex-i-quality llama.cpp-metal local code, reasoning, batch-eval 20/21 wave-core
hermes-4.3-36b ollama local reasoning, instruction-following 20/21 wave-core
glm-5.3-flash-iq3xxs llama.cpp (internal build, name withheld) batch batch-eval, synthetic-data —
granite-4.2-30b-q5km local-ollama medium chat, enterprise, grounded-answers, reasoning 21/21 wave-core
hermes-4-14b-q5km local-ollama medium chat, function-calling, structured-outputs, reasoning 20/21 wave-core
qwen3.8-flash-next-gguf-iq1s local-llamacpp fast chat, reasoning, tool-use, agentic 21/21 wave-core
mimo-v2-flash-iq2xxs local-ollama fast chat, reasoning 20/21 wave-core
mistral-small3.2 local-ollama medium chat, coding, function-calling —

source the model registry (status=live, rail=internal) · flash tier measured on wave-core-21
placement one resident heavyweight per rig — the coexistence law