A developer shares initial impressions of using a 3-bit quantized version of the Mistral Medium 3.5 model for coding tasks. The user reports that the model identified issues in their project that their previous daily driver missed, describing its performance as "really good" despite being slower than Mixture-of-Experts (MoE) models.
- Quantization: Unsloth Q3_KS with KV cache quantized to q8_0 and values to q5_0.
- Performance: Achieved 8 tokens per second with an 80k context window.
- Capability: Successfully found errors in code that the user's previous model overlooked.
The post serves as a community inquiry, asking other developers for their experiences and thoughts on using this specific dense model configuration for software development.