Groq Chat

Users are overwhelmingly impressed by raw inference speed: measured at 276 T/s for Llama 2 70B (vs ~59-60 T/s on Perplexity and Together API), and up to 460-493 T/s for Mixtral 8x7B. Multiple HackerNews commenters (including self-identified Groq engineers) emphasize that the underlying LPU hardware is a general-purpose compute fabric not designed specifically for LLMs, giving it major latency and throughput advantages for inference over GPUs, though not necessarily a compute advantage for training. Users note the chat interface uses Llama-2-chat (safety-trained), and censorship is aggressive: one user reported it refused to write a Bukowski-style poem about pumping gas, and lectured about objectification when asked for a Valentine's message. On quality, users acknowledge Llama 2 70B is 'noticeably worse than GPT-4 turbo.' One user compared GPT-4 and Groq on a Spring Boot architecture question and found them at 'about parity' except Groq suggested a specific library. Users appreciate it's free for everyday use. Several expressed desire for an API endpoint, with one speculating costs might be 'eye watering' but stating they'd switch immediately if pricing were reasonable. Technical discussions also cover integration friction — Firefox's AI chatbot sidebar doesn't work with chat.groq.com without manual configuration.

No aggregate ratings were found for this product during the last harvest.