Unbiased Bob
Deeply technical content about building and measuring coding AI on your own hardware.
LIVE ON THE RTX 4090
Videos
An Unbiased Review of Meta's Muse Spark Open
816B parameters on 2× B200s ties a 27B model on Bob-Bench Wide.
Play the skateboard games both models wrote. Transcript and ad-free torrent.
maybe wait on this one
Strata runs a 177B Qwen MoE at ~100 tok/s on one 4090. How the
expert cache works, why llama.cpp is slower, and why I won't install it.
Offline fork, transcript, and ad-free torrent.
8GB VRAM + 12GB RAM: Recommended Setup
Qwen3.6 35B-A3B on budget hardware: the config, a 1070 Ti
trial, and 11 BBW configurations. Transcript and ad-free torrent.
Not even a model
PrismML's Ternary Bonsai 2 27B against three other ways to use
an 8 GB card. Transcript, configs, and an ad-free download.
I was wrong about this
Qwen3.6 35B-A3B in 8 GB of VRAM by offloading the experts to
system RAM. Transcript, config, and an ad-free download.
16 GB VRAM / 0 GB RAM
What to actually run on a 16 GB card, entirely resident in
VRAM. Transcript, config, and an ad-free download.
Resources
-
Muse Spark Open live stream — falling sand demoA falling-sand physics sim the 816B model wrote on stream, playable in the browser, plus the prompt.
-
Bob-Bench Wide alpha — interactive paretoEvery BBW run so far: score vs. wall clock across 8, 16, and 24 GB budgets. Click a point for its full config.
Contact
admin@unbiasedbob.comTip jar ยท Bitcoin
1ub1ASpxqvMC1VPeZnEpshpeuBTMvUGKd
Helps pay for the rented GPUs.