Unsloth releases 2.5x faster Qwen3.6 NVFP4 quants with FP8 KV Cache

Unsloth has released optimized NVFP4 quantizations for the Qwen3.6 models, achieving significant speed improvements over NVIDIA's standard NVFP4 implementations without accuracy degradation.

• Qwen3.6 27B runs 2.5x faster using W4A4 tensor cores compared to NVIDIA's W4A16 approach.
• Qwen3.6 35B-A3B offers two variants: NVFP4-Fast (1.79x faster) and standard NVFP4 (1.56x faster).
• FP8 KV Cache calibration is included, enabling contexts twice as long.
• Benchmarks on MMLU-Pro, GPQA, and AIME 2025 show comparable or superior performance against NVIDIA's NVFP4, FP8, and BF16 baselines.

These updates allow users to deploy Qwen3.6 models with higher throughput and extended context windows while maintaining competitive accuracy.