GLM-5.3-Flash exlr8 fits multimodal inference on DGX Spark

This is a meaningful local-inference milestone, but “no quality cliff” needs qualification: K4 matches NVFP4 quality in reported tests, while the smaller K2 and K3 variants remain unvalidated against BF16.

• –K2 targets one GB10 box and reports up to 53 tokens per second on code workloads and 29.5 on prose with speculative decoding.
• –Multimodal support is real, though image turns may see lower speculative-decoding acceptance because the drafter uses text only.
• –K4 is 3–8% faster than the comparable NVFP4 build and enables roughly 15% more KV-cache capacity, but long-context prefill is slower.
• –The November 7 update adds mixed 2.01-bit and 3.49-bit manifests, giving developers more memory-quality tradeoffs.
• –Deployment still requires KindlingAI’s custom exlr8 stack rather than a hosted inference provider, making this most compelling for hands-on local operators.