WhichLLMModel Ranks Top LLM Text Models

whichllmmodel.com published a browser-based model ranking and local finder that evaluates which open-source text LLMs can run on a user-specified machine. The interface takes GPU VRAM, system RAM, and a desired context window in tokens as inputs, and lets users choose quantization formats and CPU offloading, according to whichllmmodel.com. The site lists ranked models and allows selecting up to four models for side-by-side comparison.

whichllmmodel.com reports per-model memory footprints for multiple quantization formats. For example, the site shows Ministral 3 14B with FP16 CPU-offloaded total 28.9 GB, Q8_0 CPU-offloaded total 15.8 GB, and Q4_K_M total 9.3 GB (weights and KV cache broken out); Gemma 4 12B appears with FP16 total 24.7 GB, Q8_0 total 13.5 GB, and Q4_K_M total 8.0 GB; Llama-3.1 8B is listed with FP16 total 17.0 GB, Q8_0 total 9.5 GB, and Q4_K_M total 5.8 GB, per the site. The tool explicitly lists compatibility for FP16/BF16, Q8_0, and Q4_K_M quantization formats and indicates when models fit in VRAM versus requiring CPU offload.

Tools that map model memory requirements to specific hardware settings help practitioners decide whether to run models locally, which quantization to use, and when to accept CPU offload trade-offs. Industry practice shows quantization and KV-cache sizing are the primary levers for moving models from multi-GPU setups down to single-GPU or CPU-offloaded configurations. The broader whichllmmodel.com site also covers cloud model comparisons with benchmark scores (swe-bench-pro, gpqa-diamond) and pricing.

This comparator is useful for ML engineers and hobbyists assessing cost and latency trade-offs for local inference. The local finder focuses on memory compatibility rather than runtime throughput, accuracy, or real-world latency - those remain separate validation steps. The site aggregates practical memory estimates for on-device deployment planning.

Monitor updates to model entries, quantization format support, and reported KV-cache sizes. Also watch for added runtime metrics (peak memory during inference, latency) or integrations with benchmark suites that would convert these compatibility checks into end-to-end deployment guidance.