● Next: meet us at Web Summit Lisbon, Lisbon, Portugal, 9–12 Nov 2026 Details →
self-hosted
intelligence
Open models · Quants · Contexts · Memory

Benchmarks for the AI you run yourself.

Pick the context you need and the memory you have, and see which open models fit your hardware and how well they actually perform.

Best overall deepseek/v4-flash-0731 unsloth-iq3_s · 109 GB at 32k
0.934
Best under 16 GB qwen/qwen3.8-27b unsloth-q4_0 · 15 GB at 32k
0.929

Round-trip translation, best of the no-reasoning and low-reasoning runs, memory at 32k context. 19 models, 47 model/quant combinations, 94 runs.