My Experience with Qwen 3.5 35B

A developer's detailed experience with Alibaba's Qwen 3.5 35B model is going viral, positioning it as a new benchmark for powerful, locally-run AI. The user tested it against other strong recent releases like NVIDIA's Nemotron Nano 30B and Zhipu AI's GLM 4.7 Flash, finding Qwen 3.5 35B smarter overall. It particularly excelled at a complex, real-world task: categorizing hundreds of confusingly-named services across three similar domains from a large homepage configuration—a job that previously required pulling out a massive 120B parameter model. Key technical specs include a massive 262,000-token context window where processing speed doesn't degrade, vision support via an mmproj file, and a fast 115 tokens/second generation speed when quantized to Q8 precision on a 48GB VRAM rig.

The review notes the model is excellent for 'vibe coding' or fluid programming assistance, though it revealed a subtle limitation: after roughly 80k tokens of context, when asked to add a line of code without explicit placement instructions, it might insert it in the wrong spot—a task where larger state-of-the-art (SOTA) models might infer the correct location. The user is now crowd-sourcing advice on whether to trade speed for quality by trying even larger quantized versions like the Qwen 3.5 122B (Q4_XS, 37 t/s) or the Qwen3 Coder 32B, and seeking experiences on using these models for agentic workflows. This real-world test underscores the rapid advancement of mid-size open-source models that are becoming viable for complex professional work.