@artyomita: Better than GPT-5.6 Sol and Kimi K3 (2.8T), and it is 27B: it runs on your own GPU. AREX-2 from BAAI scores 81.8 on MLE-bench Lite, against 72.7 for both. The loop moved into the model. Try, score, retry used to live in the agent's code. Here the model runs it: on Frontier-CS it climbs 54.4, 65.9, 70.7 over a 5 hour budget while DeepSeek-V4-Pro (1.6T) stops at 44.7 after 2 hours. A 24 GB RTX 4090 takes Q4_K_M (17.2 GB), a 16 GB RTX 4080 takes IQ4_XS (15.24 GB). Already 16 community quants (GGUF, MLX, AWQ, FP8, EXL3) a few weeks after release, and more are coming. Which model should I decode next? Code: github.com/ArtyomITA Music: "Late Night Radio" by Kevin MacLeod (incompetech.com), CC BY 4.0 #llm #ai #opensource #agents

adriandj3

Open In TikTok:

Region: IT

Monday 09 November 2026 13:00:00 GMT