AI & ML interests
Post-training, model behavior and transparent evaluation. Research from alphaBell.
Recent Activity
View all activity
Articles
A controlled GPU-kernel correctness paper and corpus, paired with alphaBell's five-card browser quiz about missed input regimes.
Eleven publisher-reported benchmark pairs, two pinned DeepSeek releases, and alphaBell’s source-checked interactive comparison.
Five measured experiments, exact CPU reproduction, real music clips and matched Flint/Qwen Open Q4_1 and Q8_0 checkpoints.
Original weights, local runtime formats and concrete debugging resources for building with Recursive Flint.
Official MOSS Base and SFT checkpoints, alphaBell spatial prompt editor, and browser-local blind layout check for one generated image pair.
-
OpenMOSS-Team/MOSS-Image-Base
Text-to-Image • 5B • Updated • 10.1k • 596 -
OpenMOSS-Team/MOSS-Image-SFT
Text-to-Image • 5B • Updated • 15.6k • 835 -
MOSS Layout Studio + Blind Check
🧩Draw MOSS layouts and blind-check two generated outputs
-
MOSS Blind Layout Check
🧪Blind A/B check of MOSS text and placement from local images
Nanbeige and Qwen publisher checkpoints plus alphaBell's source-pinned decoder map and conditional full-attention K/V calculator.
Play small games, inspect the open workshop, and study actual Flint Q8_0 generation and repair attempts.
alphaBell FP8 research checkpoint, interactive historical BF16 evaluations and the original Qwen parent.
A controlled GPU-kernel correctness paper and corpus, paired with alphaBell's five-card browser quiz about missed input regimes.
Official MOSS Base and SFT checkpoints, alphaBell spatial prompt editor, and browser-local blind layout check for one generated image pair.
-
OpenMOSS-Team/MOSS-Image-Base
Text-to-Image • 5B • Updated • 10.1k • 596 -
OpenMOSS-Team/MOSS-Image-SFT
Text-to-Image • 5B • Updated • 15.6k • 835 -
MOSS Layout Studio + Blind Check
🧩Draw MOSS layouts and blind-check two generated outputs
-
MOSS Blind Layout Check
🧪Blind A/B check of MOSS text and placement from local images
Eleven publisher-reported benchmark pairs, two pinned DeepSeek releases, and alphaBell’s source-checked interactive comparison.
Nanbeige and Qwen publisher checkpoints plus alphaBell's source-pinned decoder map and conditional full-attention K/V calculator.
Five measured experiments, exact CPU reproduction, real music clips and matched Flint/Qwen Open Q4_1 and Q8_0 checkpoints.
Play small games, inspect the open workshop, and study actual Flint Q8_0 generation and repair attempts.
Original weights, local runtime formats and concrete debugging resources for building with Recursive Flint.
alphaBell FP8 research checkpoint, interactive historical BF16 evaluations and the original Qwen parent.