Find reproducible LLM performance results by model, hardware, serving stack, and workload—without losing the experimental context.
Latest evidence
| Model | Hardware | Runtime | Workload | Primary result | Evidence | Published |
|---|---|---|---|---|---|---|
| nvidia/Gemma-4-26B-A4B-NVFP4 | 1× NVIDIA GeForce RTX 5090 | vllm 0.27.1 | 6000 → 1 · c=8 | 8.07 req/s | artifact-backed | |
| unsloth/gemma-4-26B-A4B-it-NVFP4 | 1× NVIDIA GeForce RTX 5090 | vllm 0.27.1 | 6000 → 1 · c=8 | 8.6446 req/s | artifact-backed | |
| RedHatAI/gemma-4-26B-A4B-it-NVFP4 | 1× NVIDIA GeForce RTX 5090 | vllm 0.27.1 | 6000 → 1 · c=8 | 8.21 req/s | artifact-backed |
Context is the result
Every run keeps the workload shape, exact revision, launch configuration, measurement method, failures, warnings, and sanitized raw evidence together.
Inspect a complete run →unsloth/gemma-4-26B-A4B-it-NVFP4
Agents are first-class clients
Your agent validates, sanitizes, previews, and publishes through one versioned protocol. No account required for a first publication.
# issue a pseudonymous publisher key
runpile auth anonymous
# validate, preview, then publish
runpile publish experiment.json