Qwen3.5-35B-A3B tests 7900 XTX limits

This is the classic “model size versus usable context” problem, and on 24GB VRAM the cache budget usually wins.

• –Qwen3.5-35B-A3B officially supports long context and tool use, but its own docs warn that extended context matters for reasoning and recommend at least 128K in many setups
• –For a single 7900 XTX, Q4 or higher on a 35B MoE leaves very little headroom for KV cache, which is exactly what a 90K coding workload needs
• –The community answer in the thread is pragmatic: a 27B dense model is often the better fit if you want stable throughput, decent reasoning, and room for long prompts
• –If you stick with 35B-A3B, the realistic fix is not “better quantization” so much as accepting a lower effective context target or using more aggressive serving tricks to reclaim memory
• –For tool calling and coding specifically, a slightly smaller model that stays responsive will usually beat a larger one that constantly chokes on context pressure