deepseek-v4.1VIDEO
DeepSeek V4.1 Flash vs Nanbeige 4.3 33B on a Doom-building task
On a Python Doom-style game rendered in tmux Braille, DeepSeek V4.1 Flash 1017 at ~40 tok/s on 2×3090 kept producing results while Nanbeige4.3-33B at 70 tok/s on one 5090 overthought.
Aleksey Smolenchuk
@lxeArchived 2 days agoOriginally posted 2026.10.26
Aleksey Smolenchuk @lxeOriginally posted 2026.10.26
Putting Nanbeige4.3-33B through the Doom gauntlet on a single RTX 5090. 2.5 bpw EXL3, bespoke Rust + CUDA runner: 70 tok/s Against DeepSeek-V4.1-Flash 1017 on 2× RTX 3090 at ~40 tok/s. The task: build and run a playable Doom-style game in Python, rendered with colored Braille in tmux. DeepSeek consistently produces results while Nanbeige consistently overthinks. I'm validating it against original 16-bit but so far it tends to overthink as well :/
x.com/lxe/status/2114458953329639485Archived 2026.10.26 12:02