LM Evaluation Harness

GuideLLM is an open-source platform for benchmarking and evaluating real-world LLM inference deployments. It simulates end-to-end interactions with OpenAI-compatible and vLLM-native servers, using real or synthetic text and multimodal datasets and configurable execution profiles such as synchronous, concurrent, throughput, constant-rate, Poisson, and sweep workloads. It measures SLO-oriented latency and token statistics, including time to first token, inter-token latency, and end-to-end behavior, then produces console, JSON, CSV, or HTML reports for deployment tuning, capacity planning, and regression tracking. The tool provides both a CLI and API, with multiprocessing, threading, asynchronous execution, and Hugging Face, file, synthetic, or custom data sources.

llmfit is a terminal tool that evaluates which local language models fit and run well on a user's hardware. It detects system specifications including RAM, CPU, GPU, and multi-GPU configurations, then scores models across fit, estimated speed, quality, and context dimensions while accounting for quantization and mixture-of-experts architectures. It provides an interactive terminal UI and a classic CLI, supports local runtime providers including Ollama, llama.cpp, MLX, Docker Model Runner, and LM Studio, and can output recommendations as JSON. Its benchmarking workflow can measure real token-per-second performance on a machine and contribute those results to the project's fit data.

Langfuse is an open-source LLM engineering platform for developing, monitoring, evaluating, and debugging AI applications. It ingests traces for LLM calls and related operations such as retrieval, embeddings, and agent actions, with integrations including OpenTelemetry, LangChain, the OpenAI SDK, LiteLLM, and other LLM frameworks. Its features include prompt management with version control and collaborative iteration, LLM-as-a-judge and code-based evaluations, user feedback and manual labeling, datasets for test sets and benchmarks, trace replay and debugging, metrics, and a playground for testing prompts and model configurations. Langfuse exposes APIs and typed Python and JavaScript/TypeScript SDKs for custom LLMOps workflows. It is available as a managed cloud service or for self-hosting with Docker Compose, virtual machines, Kubernetes via Helm, or Terraform templates.

FrontierHarness Eval is an open evaluation benchmark and repository for measuring how coding-agent harness configurations affect software-engineering task results when the underlying model and runtime are held constant. It publishes benchmark definitions, task instructions and metadata, harness-version metadata, normalized aggregate and task-level results, and an agent-neutral workflow for evaluating additional harnesses. The benchmark runs the same model through multiple harness configurations on 30 tasks, using a frozen golden checkpoint and a fresh restore for each task. Its scripts provision the environment, run trials, record verifier-based pass/fail outcomes, measure cost, cache behavior, and speed, then normalize results and generate charts and reports. The repository also provides a CLI and skill-based workflow for evaluating a third-party harness, with infrastructure failures marked separately from task failures. The project focuses on terminal-based software-engineering tasks and states that its results may not generalize to other knowledge-work domains. The repository is published under the frontier-harness-eval organization, with isolated runtimes and checkpoint restores provided by Runta.

vLLM is an open-source library and inference-serving engine for deploying large language models on GPUs and other hardware accelerators. Originally developed in UC Berkeley's Sky Computing Lab, it is maintained by a broad community and provides an OpenAI-compatible API server, as well as Anthropic Messages API and gRPC support. Its serving stack manages attention key-value memory with PagedAttention, continuous batching, chunked prefill, prefix caching, streaming generation, and disaggregated prefill, decode, and encode. It also supports speculative decoding, quantization, optimized attention and GEMM/MoE kernels, automatic kernel generation and graph-level transformations, structured outputs, tool calling, multiple decoding algorithms, and distributed inference through tensor, pipeline, data, expert, and context parallelism. The project integrates with Hugging Face models and supports decoder-only, mixture-of-experts, hybrid state-space, multimodal, embedding, retrieval, reward, and classification models. It can run across NVIDIA, AMD, Intel, and other supported accelerators, as well as x86, ARM, and PowerPC CPUs, and can be installed with uv or pip or built from source.