Vibe coding with with Llama 4 on Groq Inc is on another level in terms of speed! Unfortunately, quality of the output and instruction following capabilities of Scout is not there yet. Will try Maverick soon!

Tried Llama 4 Maverick also on Cline (still not available on Groq Inc). It is much better in terms of instruction following and the output quality but still it doesn't feel on the same level with Sonnet 3.5v2/3.7/DeepSeek 3.1/Gemini 2.5 Pro/Grok 3 on coding.

Tested Llama 4 Scout - 6 bit quantized version offline on M4 Max and it is fast! Got ~30t/s which is faster than most of the hosted models.

Tested Llama 4 Scout 4-bit quantized version also on MLX and it is running at ~40t/s and consumes around ~61GB of memory. It seems 64GB Macs won't be enough to run this without spilling into swap.

I have been playing with Qwen3 series models and they are very powerful in agentic use cases. With reasoning capabilities, even the small 4B model can handle several tool calls over MCP properly and give the right answer accordingly! Very impressive.