Not available in English yet
4B 模型 23 步挖到铁矿:不写字的 Agent 在 RTX 3090 上跑通了
Related Reading
More on #Mica
A $30-Trained 4B Model Beats Rivals 4x — "Bigger Is Better" Is Cracking
Indie dev trained 4B Mica for $30. In Tetris, it cleared 4x more lines than a peer model—hinting small specialists may undercut "scale is everything."
27B Model Codes Overnight at Home — But Managing the Agent Is the Real Question
Jan ran Qwen3 27B on two 16GB GPUs, coding overnight. Real story: file-based memory, step verification, human veto. AI coding has gone from big-budget
AI Customer Service Tutorial Hits 25 Chapters: Time to Recalculate Costs
25-chapter tutorial modularizes AI customer service into 18 steps. SMEs who spent millions outsourcing should now recalculate true engineering costs.
DeepSeek's Three-Move Combo Crashes Agent Costs — LLM Era Shifts to Order War
DeepSeek's three moves crashed Agent costs; China's LLM race shifts to orders. Kling, WPS 365 monetize AI; Kimi's drop, Manus's failure warn caution.
50 Records In, 20 Seen by Model: Agent Silent Truncation Is the Real Killer
A tech team found AI agents read 50 database records but silently passed only 20 to the model with no errors raised. Traditional "no error = success"
Agent Projects Don't Fail Because Models Are Dumb — They Fail After HTTP 200
An engineer recaps an Agent failure chain on Juejin: errors collapsed into one code, but real crashes scatter across 7–8 stages, mostly after HTTP 200