BitNet 1.58-bit support is a different quantisation regime rather than simply a smaller integer, and a separate compute-in-memory block shifts the pitch from peak throughput to running a model continuously at very low power. A 128K on-device context is a step change over the previous generation. Treat the circulating 100 TOPS figure with care: it is attributed to MediaTek by a third party but appears nowhere in MediaTek’s own press release, product page, blog or launch infographic, and no one states its precision.
MediaTek Dimensity 9500
Two genuine firsts: hardware support for 1.58-bit ternary models, and a compute-in-memory NPU for always-on work.
Silicon from MediaTek
4 chips on record, newest first.
-
2026
MediaTek Dimensity 9600 Pro
2nm NPU 1090, second-gen Super Efficient NPU
the NPU 1090 delivers 51% higher LLM prefill performance, 55% higher Token generation per watt
Relative only -
2025
MediaTek Dimensity 9500
TSMC N3P NPU 990 with Generative AI Engine 2.0, plus a separate compute-in-memory always-on NPU
2x compute; 100% faster 3B-parameter LLM output; 128K-token on-device context; BitNet 1.58-bit support. The widely cited 100 TOPS figure is not MediaTek’s; it traces to Counterpoint Research
Relative only -
2024
MediaTek Dimensity 9400
Second-generation TSMC 3nm NPU 890
Up to 80% faster LLM prompt performance and up to 35% more power efficient than Dimensity 9300
Relative only -
2023
MediaTek Dimensity 9300
Third-generation TSMC 4nm APU 790
8x faster than the previous generation for Transformer models; supports LLMs at 1B, 7B and 13B parameters, scalable to 33B
Relative only