Hi Groq Community! 
I am building an open-source, news-aware Reinforcement Learning (RL) agent specialized for the Indian Nifty 500 market.
My architecture uses a “Teacher-Student Distillation” workflow:
-
Teacher: Llama 3.3 70B (Versatile) on Groq provides a “Ground Truth” expert analysis for 479 stocks based on technical chart patterns and live news sentiment.
-
Student: A lightweight Qwen-0.8B model is being fine-tuned to mirror these expert rationales for ultra-low latency inference.
The Challenge: Your LPU speeds are incredible (300+ TPS), which is exactly what this agentic flow needs. However, I’ve hit the 100K Token-Per-Day (TPD) limit on the Free Developer Tier. To generate a high-quality historical dataset for all 479 stocks, I need to process roughly 5M+ tokens.
Does anyone have advice for:
-
Pacing massive batch-inference runs for “burst” dataset generation?
-
The current status of the Developer Tier upgrade process (mine is showing as closed)?
-
Whether there is a program for hackathon participants looking for temporary limit increases for research purposes?
You can check out the project progress here: [https://github.com/harshgmx47/openenv-nifty500-stock-agent-rl\]
Really enjoying the Llama 3.3 70B performance so far—the latency is unmatched!
#FinanceAI #Llama3 #GroqSpeed #Nifty500 #ReinforcementLearning