Robot sim benchmarks have a physics problem. An audit of RoboDojo found loose joints and ignored material properties, and the maintainer agreed.
If the sim physics is wrong, what did the leaderboard measure?
1
37
A 9B model at 59 tok/s on a 16 GB GPU.
JJ Eaton @jayleaton ran SenseNova-Loop-9B on TensorFold: 59.1 tok/s on an RTX 5070 Ti, 59.5 on an M5 Ultra.
Weak spot he flagged himself: tool calls, 58% on BFCL.
Fine for chat on a cheap card. Not your agent brain yet.
DeepSeek-V4.1-Flash at 59.5 tok/s on about 160 W.
That's one M5 Ultra 512 GB, measured by Justin Kalland and shared by @OwnYourCompute.
At home the power is the small part. The box is the real cost.
The new RTX Spark desktops use the same Grace CPU and Blackwell GPU class as DGX Spark, but ship 10GbE instead of ConnectX-7. No 200G link, so no two-box setups for the big MoEs.
We run local inference on DGX Spark. If you might cluster later, buy the box with the 200G port.
3
2
141
Same RTX 5090, same small model, 4x the throughput.
@flaviusburca tested 8 users at once: Surogate served 2,765 tok/s, llama.cpp 683.
Single-user benchmarks hide this. Bench at the user count you really run.
2
1
117
Hy4 Flash read a 1M-token prompt in 44 min on a Mac Studio M3 Ultra and 81 min on a DGX Spark, per @Razumasu.
Follow-ups on the cache took 2.8 s on the Mac, 3.4 on the Spark. Slow read once per doc.
Which doc would you pre-read overnight?
1
99
Four DGX Sparks, one week of tuning, and code went from 26.85 to 56.71 tok/s.
@arlogilbert also cut the slowest first token at 8 users from 64 s to about 1 s by raising the sequence cap.
His biggest wins were a depth-3 drafter, FP8 attention (+12%) and a faster path for small all-reduces (+16%).
Tune the four boxes you have before you buy a fifth.
1
93
OpenAI's coding agents now finish 46% of 2 to 4 hour research tasks with zero human help. In March it was 23%.
By task, it's 79% on monitoring training runs and 25% on writing inference and serving code.
Agents are good at watching. Still weak at building.
Which job would you hand to an agent first?
2
82
Numbers from OpenAI's own write-up, tasks run Aug 1 to Oct 18: openai.com/index/research-ac…
1
18
Looking to buy one to use here in Thailand when shit will hit the fan big.
Plug straight into the sun and let the DJI Power Auro 2000 Elite fast-charge your entire ecosystem in the wild. Unfold the panels anywhere and turn sunlight into uninterrupted flight time. Pure off-grid power for endless creative sessions.
4
130






