NVIDIA RTX 5090
- Memory
- 32 GB GDDR7
- Job
- Fast local inference, evaluation runs, and the agent loop. Runs a 27B model at Q4_K_M with room for long context.
- Measured
- About 100 tok/s on Qwopus 3.6 Fusion with MTP speculative decoding
The lab
Everything I publish comes out of this room: one RTX 5090 and one NVIDIA GB10, paid for out of my own pocket and run on my own time. Since May, the GB10 has carried most of the compute for the Qwopus line. I'm still surprised by how much two devices can do.
Hardware
NVIDIA GB10
- Memory
- 128 GB unified
- Job
- Training. Since May 2026 it has carried most of the compute for the Qwopus line on Jackrong's repositories.
- Why it matters
- Enough memory to fine-tune 27B-class models at home, on a device that sits on a desk
Entirely self-funded: the hardware, the infrastructure, and the token spend. The work happens in my free time.
Working with LLMs in the loop
I've built with language models since GPT-3, and they're a real part of how one person ships this much. Growing up alongside them taught me the most useful skill in this field right now: knowing exactly where they fail.
I design agent loops on purpose, check the output against production standards, and don't mistake fluent for correct. My own finetunes run in that loop too, for long experimental runs that wouldn't justify an API bill, and for when the API is down.
Some of my best results came from experiments frontier models told me not to try. GLM-18B is one of them.
Design the loop
Decide what the agent may do, what it must prove, and where a human checks.
Run it
Frontier APIs where they earn their cost. My own local finetunes for long, uncertain runs.
Verify
Tests, hidden verifiers, and review to production standard. Fluent output is not evidence.
Ship or discard
What survives goes out. What doesn't becomes data for the next run.