Search

GitHub - ENTERPILOT/ai-gateway-reproducible-benchmark: Comparing Bifrost vs LiteLLM vs GoModel vs PortkeyAI. AI gateway benchmark. · GitHub

A reproducible benchmark suite compares multiple AI gateways running on AWS, recording gateway overhead across several versions and providing detailed, timestamped results and charts. It includes setup, execution, and teardown steps, with configurable environments and workloads.

GitHub - deepu105/quant-bench: Speed and accuracy benchmark for local LLM quants across engines, with repeated runs and best/worst per run · GitHub

A Python-based benchmark toolkit to measure speed and accuracy of local LLM quantizations across multiple engines, using repeated runs and fixed seeds to report mean, best, and worst results. It includes setup instructions, problem sets, server management, and outputs (summary and per-run data).

qwen38-megakernel/docs/BENCHMARKS.md at main · cutlerbenjamin1-cmd/qwen38-megakernel · GitHub

The document details benchmark methodology and results for a Qwen-based model on a fixed hardware setup, emphasizing session-to-session variability, caching effects, and cross-build comparisons with specific commits and configurations. It documents timing metrics, context sizes, and decoding strategies used across multiple runs.

Qwen3.8-27B-FP8 · 4 B65 · FP8 | Intel Arc Pro B65 inference results

The document presents benchmark metrics for a specific AI model, detailing throughput and latency across various context lengths and request loads, with notes on queueing and capacity limits. It clarifies measurement conditions, including prefill, warm-up, and how to interpret over-capacity and queued requests.

[2610.27234] EGI-BENCH: An Enterprise General Intelligence Benchmark Suite

EGI-BENCH unifies eight enterprise agent benchmarks behind a single CLI, enabling a six-dimension EGI score derived from sub-metrics, with results for fourteen frontier and cost-efficient LLMs across GPT, Claude, and Gemini in quick mode.

LegoBench - Porsche 911 GT3 RS from instructions — Gemini 4 Argon (High) build

Discusses multiple benchmark builds, comparing AI models and setups by reference accuracy, with the current build omitted from the list.

GitHub - blawesom/gpu_inference_bench: nvidia/amd/intel comparison test · GitHub

This repository benchmarks vLLM inference on NVIDIA, AMD, and Intel GPUs using Docker, configurable model and optimization tests, and per-machine reports. It also offers a quick test and a GPU-memory fit check.

rackbites — build a used server from live eBay prices

Rackbites uses live eBay listings to price used server and GPU builds, with shipping included and very low GPU asks filtered out. It also lists language-model benchmarks and links to the sources, setup details, and buying checks.

Claude Fable 5.2 and Mythos 5.2: Benchmarks, Pricing, Access

Translated title. Original: Claude Fable 5.2 und Mythos 5.2: Benchmarks, Preise, Zugang

Anthropic has released Claude Fable 5.2, reporting strong results across most of its benchmark comparisons and adding tools to guide or delegate work to smaller models. The article also describes its safeguards, restricted access to Mythos 5.2, and limitations in the benchmark evidence.

FrontierHarness Eval: live AI coding agent benchmark run

FrontierHarness Eval compares 21 AI coding agent harness configurations across five models and 62 tasks, measuring pass rate, cost, and speed. It scores verifier-approved attempts and tracks infrastructure failures separately.