Gemma-4-26B-A4B hits 15 tps via Unsloth

The arrival of MXFP4 quantization for the Gemma 4 family marks a significant leap in making high-quality Mixture-of-Experts (MoE) models accessible on budget hardware.

• –MoE architecture with 4B active parameters delivers 26B-level reasoning with the compute footprint of a much smaller model
• –MXFP4 (Microscaling Floating Point) maintains higher precision than traditional 4-bit integer quantization, preserving model intelligence
• –Achieving 15 tps on a sub-$200 GPU like the RX 6600 proves that sophisticated AI is no longer gated by high-end VRAM
• –Support for 256K context windows via Unsloth optimizations enables large-scale local RAG and document analysis
• –First-class Vulkan support in LM Studio further expands the hardware compatibility beyond just NVIDIA users