870 MB
TQ4 used 870 MB. The bf16 comparison used 2.7 GB.
Compact models and tools for AI agents.
Torad's research asks how much model capability fits on hardware one person can own. Its tools show you what your coding agents are doing and give every AI tool you use one shared memory.
- Temple
A live, agent-drawn picture of any codebase, shared by you and the agents writing it. Invite-only beta on Linux x64.
- Brain
One memory that every AI tool you use shares.
- Model research
TQ4, QGRE, native FP4, and a small training-and-inference megakernel.
Compact-model research
Current Torad runs measure memory, throughput, and training on consumer hardware. Independent reproduction is still pending.
238 tok/s
vLLM reached 267 tok/s in the comparison.
13.5 GB
Peak use on an RTX 5080 with 16 GB.