70-80B coding models still dominate for front-end work despite smaller rivals
Reddit user with 3x24GB VRAM asks: which 70-80B model is best today?
A Reddit user with three 24GB VRAM GPUs is on the hunt for the best recent coding model in the 70-80B parameter range. They currently use Qwen-coder-next but are open to alternatives. The key requirements: support for a Q6 quantization, at least 256K context (critical for coding), and recent training data—especially for rapidly evolving front-end frameworks. While they acknowledge speed trade-offs, they believe larger models still outperform dense 27-31B models in complex reasoning tasks. They also note that micromanaging AI agents yields better results than letting them run autonomously, even if it's slower.
The discussion highlights a persistent tension in local AI: smaller models offer speed but lack depth, while bigger models demand more VRAM and slower inference. The user's setup (3x24GB) is common among enthusiasts, and their focus on front-end work reflects a real-world need for models that stay current with JavaScript/TypeScript frameworks. The post underscores that for professional coding tasks, a 70-80B model at Q6 quant still provides a sweet spot between capability and hardware constraints, especially when context length is non-negotiable.
- User runs 3x24GB VRAM GPUs and targets 70-80B parameter models at Q6 quantization.
- Requires 256K context minimum for coding, dismissing smaller 27-31B dense models as inferior.
- Qwen-coder-next is a current fallback but user wants newer models trained on recent front-end data.
Why It Matters
Highlights the ongoing demand for mid-size local coding models that balance context, quantization, and up-to-date training data.