ok
20B on a 16GB GPU
GPT-OSS 20B, MXFP4, llama.cpp, on the RTX 5060 Ti. Generated tokens per second.
95.72tok/sMeasured
Kept
Wins
Things that actually happened on the IronClaw machine. One measured outcome each. No before-and-after pair I made up.
ok
35B actually ran locally
Qwen3.5-35B-A3B on 20 June 2026, on the RTX 5060 Ti. About 3B active parameters, mixture of experts, so not a dense 35B.
~64 to 68tok/sMeasured
ok
The security gateway held
AgentGate rejected an unknown tool request. That is the response code.
403HTTPMeasured
ok
Long-context testing worked
GPT-OSS processed a prompt of about 120K tokens and generated at this speed.
61.3tok/sMeasured
ok
The house started answering
Home Assistant had every approved entity state available after the October 2026 deployment.
34of 34Measured