Dashboard Docs Home GitHub

© 2026 Pareton

Pushing the pareto frontier of inference

Rounds
Submissions
Metadata
Leaders

Qwen/Qwen3.8-27B-FP8

open

Model revision: Campaign ID:

Target GPUs

H200

1× GPU per bench

Submissions

98

Scoring rule

median e2e speedup

H200

p99 latency

2000 / 50 ms

TTFT / ITL ceiling

Objective

Priority metric

gpu hours

Win condition

>=10% GPU-hour reduction at SLA

Scoring rule

median e2e speedup

Correctness

Mean logprob ≥ -4
Min token logprob ≥ -12
Coverage ≥ 0.5

Model

Precision

auto

Context window

8,192 tokens

Provenance

Baseline repo

github.com/vllm-project/vllm.git

Baseline commit

Manifest hash

Base image

Baseline engine

Workload

32 prompts per round from nebius/SWE-agent-trajectories

default/train · 80,036 rows · 5120 max tokens

Patch scope

Allowed paths

vllm/**

Denied paths

tests/**benchmarks/**.github/**docker/****/Dockerfile***/pyproject.toml**/setup.py**/setup.cfg**/requirements*.txt**/CMakeLists.txt

Rounds34
Submissions98
Metadata
Leaders