ax@ax-radar:~/curated $ grep -l 'curated=true' sources/
33 srcsignal 72%cycle 04:32

curated · 2026-08-19

4 items · updated 3m ago
2026-08-19 · Wed
20:42
34d ago
AI HOT (Curated Pool)· aihot-apiZH20:42 · 08·19
FastMetal generates 5-second 480P video on Mac in 30 seconds, no CUDA or cloud needed
FastMetal brings the FastWan-QAD family to Apple Silicon. A 5-second 480P clip runs entirely on a Mac in 30 seconds, using 3.9 GiB of memory. The DiT, DMD sampler, and decoder all run via MLX on Metal with default INT8 quantization. Three model sizes: 1.3B for 480P, 5B for 720P, and 14B targeting higher quality. The post doesn't disclose generation speed or memory usage for the 14B model.
#FastMetal#FastWan-QAD#Apple Silicon
editor take
5s 480P clip in 30s on a Mac at 3.9 GiB—useful, but the post skips 14B speed and quality comparisons.
HKR breakdown
hook knowledge resonance
open source
72
SCORE
H1·K1·R0
13:48
34d ago
AI HOT (Curated Pool)· aihot-apiZH13:48 · 08·19
Liquid AI drops 4 quantized checkpoints, recovers 97% of lost accuracy
Liquid AI released four Q4_0 GGUF checkpoints (230M–2.6B) trained with quantization-aware distillation (QAD). They keep native Q4_0 memory and speed while recovering 97% of BF16 average accuracy loss. The post doesn't disclose benchmark scores or downstream task results.
#Liquid AI
editor take
Liquid AI claims QAD recovers 97% of BF16 accuracy on Q4_0 quantized 230M–2.6B models, but no benchmark scores are disclosed — I'd discount this a bit.
HKR breakdown
hook knowledge resonance
open source
62
SCORE
H0·K1·R0
01:03
35d ago
AI HOT (Curated Pool)· aihot-apiZH01:03 · 08·19
GLM-5.3 Launches: AA Intelligence Index 60 Points, Ties for Open-Source First, Lower Cost
Zhipu released GLM-5.3, scoring 60 on the AA Intelligence Index, tying for first among open-source models. The post was blocked by WeChat, so it doesn't disclose evaluation details, parameter count, or pricing. The title only mentions 'lower cost' without specific numbers.
#智谱#Zhipu
editor take
GLM-5.3 ties for first on the AA Intelligence Index among open-source models, but the post is blocked by WeChat—no numbers on cost or benchmarks yet.
HKR breakdown
hook knowledge resonance
open source
39
SCORE
H0·K0·R0

more

feeds

admin