MediaTek's 2nm Dimensity 9600Pro runs 30B MoE models on-device with rebuilt dual NPUs
量子位 · wechat · 2026-09-16
MediaTek unveiled the Dimensity 9600Pro, a 2nm flagship chip with 33 billion transistors and an AI-native fused architecture that can run 30B-parameter MoE models on-device.
- AI: The dual-NPU design was rebuilt — the efficiency NPU cuts power 40% for background sensing tasks, while the performance NPU delivers 50% more compute per watt for LLM inference. Zurich AI Benchmark score tops 25,000, prefill speed is up 60%, and smart memory management speeds up model cold starts by 27% while keeping 25 apps alive. The Agentic Engine 3.0 promises two years of software updates to keep pace with model evolution.
- CPU/GPU: 2+3+3 all-big-core design with 34.5MB cache; Geekbench 6.4 scores near 4,300 single-core (+17%) and 12,600+ multi-core (+15%). Big-core power drops 37%, multi-core 61%. The Arm G715-based GPU adds neural rendering for real-time 720P-to-1080P upscaling, with ray tracing up 18%.
- Imaging: A joint ISP+NPU architecture enables all-in-focus group photos, 60fps focus tracking, 4K60 video, and 240fps slow motion.
The vivo X500Pro series launches Monday as the first phone to use it, followed by the OPPO Find X10 Pro Max on Tuesday.
Related event: MediaTek Unveils 2nm Dimensity 9600 Pro Flagship Chip(3 posts)→
More from Infra
- Transformers.js 4.3 ships structured output, new model architectures, and Safari WebGPU support — nicodotdev · 2026-09-16
- Novita open-sources Chord W4A16 MoE kernels, up to 2.15x faster Kimi K2.x inference on B300 — vllm_project · 2026-09-16
- VCs say the GPU shortage is really a capital access problem: 30% down payment prints money — sarahdrinkwater · 2026-09-16
- Developer builds Linux subsystem for Haiku with OCI runtime and container support — unixterminal · 2026-09-16
- Anthropic signs lease at Queensland data center park costing ~$30B, online 2027 — DigitalColmer · 2026-09-16
- Transformers models now run natively in vLLM with no port required — pcuenq · 2026-09-16