Qwen3.5 122B stumbles at 100K

Hot take: this smells like a runtime or quantization problem, not the base model suddenly running out of context headroom.

• –The official model card says Qwen3.5-122B-A10B supports 262,144 native tokens and can be stretched further with RoPE scaling, so 100K should still be inside its design envelope.
• –The olka-fi MXFP4 pack is a third-party quant; its own card shows conservative vLLM guidance and only quantizes the expert MLP weights, so calibration or inference behavior is the likely weak point.
• –The Reddit thread already has contradictory reports, including users saying NVFP4 or other setups do not reproduce the collapse, which points to stack-specific behavior.
• –For anyone evaluating Qwen3.5 locally, this is a good reminder to test the exact model, quant, and serving engine combination, not just the base checkpoint.