Advertising disclosure: we earn commissions when you shop through the links below.
Dynabook's 16-inch Copilot+ laptop, released in Japan on 2026-11-20. It pairs an Intel Core Ultra 7 355 (8 cores, up to 4.7GHz) with a 49 TOPS NPU, 16GB LPDDR5X memory and a 512GB PCIe 4.0 SSD. The 16:10 WUXGA non-glare screen, Wi-Fi 7 and two Thunderbolt 4 ports target everyday office and AI-assisted work, with up to 12 hours of video playback. Reported price is $1,990.
Manufacturer's suggested retail price. Current prices can be higher or lower. This is not a live price.
The dynabook C9 enters the market as a lightweight 16-inch productivity laptop built around Intel Core Ultra Series 3 architecture, specifically the Core Ultra 7 355 processor. Originally introduced to the Japanese market in late 2026, the laptop adheres strictly to Microsoft Copilot+ PC requirements, pairing an integrated neural processing unit (NPU) capable of up to 49 TOPS with 16GB of dual-channel LPDDR5X-7467 memory. Encased in a 1.79 kg chassis with a 16:10 non-glare WUXGA screen, Wi-Fi 7, and dual Thunderbolt 4 ports, the device targets everyday enterprise workflows and mobile knowledge workers.
Evaluating Unknown hardware for AI development or boutique regional releases requires looking past surface-level marketing claims. While Copilot+ certification ensures compliance with Windows-level AI functions like live translation, local studio effects, and semantic indexing, assessing the dynabook C9 for AI workloads requires examining its memory subsystem, unified memory allocation, and sustained inference speeds. At an MSRP of $1,990, it operates in the prosumer thin-and-light bracket, competing directly against business ultrabooks powered by Apple Silicon or AMD Strix Point processors.
For AI engineers and local inference enthusiasts, this machine is not a replacement for a dedicated workstation GPU or high-density rackmount cluster. Instead, it serves as a portable, power-efficient testing ground for edge deployment, quantized small language models (SLMs), and client-side agent integrations that leverage OpenVINO or DirectML.
Understanding the dynabook C9 AI inference performance starts with the silicon pipeline. The Intel Core Ultra 7 355 features an 8-core, 8-thread architecture running up to 4.7GHz alongside an integrated NPU delivering 49 TOPS (trillion operations per second) at INT8 precision. While the NPU meets Copilot+ standards, raw TOPS metrics tell only part of the story. In real-world local LLM execution, memory bandwidth and addressable capacity dictate operational limits far more than raw theoretical compute.
The primary system specifications directly impacting AI tasks include:
Memory bandwidth is the decisive factor in autoregressive token generation. At roughly 119.5 GB/s, the LPDDR5X-7467 setup outperforms standard DDR5 laptop memory (which typically hovers around 70 to 80 GB/s), enabling acceptable prompt processing and generation latency for small parameter sizes. However, dynabook C9 VRAM for large language models is constrained by the architecture of Windows 11 unified graphics memory. Windows dynamically partitions the 16GB system pool, capping dedicated video memory for the integrated GPU at approximately 9.2GB.
When accounting for baseline operating system overhead (usually 3.5GB to 4.5GB under standard workloads), available workspace for loading model weights, KV caches, and runtime contexts settles around 8GB to 9GB. This strict ceiling defines the boundaries of what this laptop can run without swapping heavily to the 512GB PCIe 4.0 SSD, which drastically collapses token throughput.
Power efficiency remains a core strength. The system draws roughly 10W during standard execution and caps out at a 65W thermal design envelope. This allows sustained local processing without aggressive thermal throttling or fan noise, making it suitable for quiet desk environments or mobile use where battery longevity matters.
Anyone seeking hardware for running 70B parameter models must look elsewhere: a 70B model quantized to 4-bit requires more than 38GB of dedicated VRAM, exceeding this laptop's entire hardware memory pool by more than double. The dynabook C9 local LLM footprint is instead squarely optimized for 1B to 8B parameter models.
Compact models represent the sweet spot for the integrated NPU and GPU via runtimes like OpenVINO or ONNX:
Standard open-weight workhorses run reliably under 4-bit quantization (Q4_K_M or INT4), provided context lengths remain constrained:
The optimal configuration on this machine is INT4 or Q4_K_M quantization. Stepping up to Q8_0 or FP16 on 7B models pushes memory consumption past 8GB, risking out-of-memory errors or aggressive system disk paging.
Lightweight multimodal models such as Moondream2 or SmolVLM run fluidly for basic visual classification and visual question-answering. However, running larger vision-language models like LLaVA 1.6 7B will severely limit the remaining context length, restricting long-form multimodal analysis.
The dynabook C9 targets a specific cross-section of professionals who need local AI capabilities integrated into an enterprise-ready Windows machine:
This machine is not suited for fine-tuning, LoRA training, or acting as an inference server for multiple concurrent users. Soldered 16GB memory and a single 512GB SSD severely limit deep-learning research pipelines.
When evaluating the best hardware for running AI models locally in the thin-and-light laptop category, two alternatives serve as direct comparisons:
The 15-inch MacBook Air sits in a similar price bracket. Apple unified memory architecture allows the system to allocate a larger percentage of its 16GB RAM directly to Metal compute compared to Windows DirectML limits. The M3 also delivers slightly higher sustained memory bandwidth (around 100 to 150 GB/s depending on load) and exceptional token-per-second performance across llama.cpp via Metal. However, the MacBook Air lacks active cooling, throttling sustained workloads over extended inference sessions. The dynabook C9 provides active cooling, dual Thunderbolt 4 connections, native Gigabit Ethernet, and an NPU natively supported by Windows 11 Copilot+ software.
Laptops built on AMD Strix Point (such as the Ryzen AI 9 365 or HX 370) feature competing XDNA 2 NPUs delivering up to 50 TOPS, paired with Radeon 890M or 880M integrated graphics. AMD RDNA 3.5 graphics generally hold an edge in raw compute throughput for OpenCL/Vulkan workloads over Intel integrated graphics. Furthermore, ThinkPad chassis configurations often allow memory upgrades up to 32GB or 64GB. The dynabook C9 is strictly locked at 16GB soldered memory, meaning users who require larger context windows or concurrent model execution will find better longevity in modular AMD alternatives.
The dynabook C9 stands as a purpose-built, highly portable business machine. While not the best AI chip for local deployment if raw model capacity is your primary metric, it effectively operationalizes sub-8B LLM inference and Copilot+ NPU acceleration within a lightweight 16-inch corporate form factor.
The top models this device can run at 4-bit, ranked by fit and speed.
Specs not available for scoring. This product is missing VRAM or memory bandwidth data.

Also in Our Store
Docks, fast storage, cables, and a UPS sized for a local AI machine.

Rent this class of GPU by the hour before you buy. Live prices from RunPod and Vast.ai.

Find the break-even between buying this hardware and paying for a cloud API.
Mac vs NVIDIA for local inference, if you are still choosing a platform.