"Weights & Biases" matching MCP servers:

Checks whether an open-weight LLM fits your GPU: VRAM for weights, KV cache and overhead at each quantisation, a tokens-per-second ceiling, the longest context that fits, and what would work instead when it doesn't. Reads any Hugging Face repo's config.json. Free, read-only, no API key.

Enables zero-dependency INT8 symmetric quantization of LLM weights and activations on a per-tensor or per-channel basis, reporting SNR and MSE reconstruction telemetry along with TTFT/TPOT latency and jitter metrics. Also exposes edge inference primitives such as PagedAttention block allocation, radix prefix caching, and speculative decoding verification through the Model Context Protocol.