The battle for true RDMA between MLX and CUDA continues!
Cloud models refuse to help, and I don't have 4 Sparks to cluster together to run the likes of GLM 5.2. My only other option is to spend my time on the layer around the model. Over the past 6 months, I've curated a
BRO, Kimi K3 has just been released and the effect is immediately felt in the market.
🤯📉 The same week, several AI companies whose models were considered competitors also came under pressure: 📉 → had dropped ~30% 📉 MiniMax → ~16% 📉 Alibaba → ~4% Meanwhile, the US lab said
This is f**king gold
DeepSeek just dropped V4 Pro 0813 and open sourced its agent framework the same day 😳 what you get: - DeepSeek V4 Pro 0813 - 1M context window - agentic reasoning - open weights under a permissive license - DeepSeek Harness v0.1 - MIT licensed agent
Deepseek V4 vs Qwen 3.8
An interesting side by side that tells you the strengths of both models. Qwen (vision) looks impressive at a glance but took many iterations and kept reopening to check its work. Deepseek one-shot the code blind (no vision) Looking closer: only
You have reached the end of the archive
All of deepseek