StorageReview’s ThinkPad P1 Gen 9 test delivered 5,083 points in Cinebench 2026 multi-core and ran two small-model inference tests, but its RTX PRO 2000 had 7.7GB of GPU-usable memory, below the roughly 16GB required for the review’s larger model-load tiers. Those results make the tested configuration a stronger fit for CPU-heavy work and the specific small-model tasks that ran than for the larger local-AI tiers in that test suite.
The configuration behind the results
StorageReview tested SKU 21UE000XUS with an Intel Core Ultra 9 386H, 64GB LPCAMM2 memory, a 2TB PCIe 5.0 SSD, an NVIDIA RTX PRO 2000 Blackwell GPU with 8GB of memory at 50W, and a 16-inch 3840 × 2400 IPS display. The laptop was plugged in, with Windows set to Best performance and Lenovo’s Ultra Performance mode enabled.
The video review by Tech Deals describes a different P1 Gen 9 configuration: a Core Ultra X7 368H, 32GB of RAM, and a 1TB SSD. Its system is not the 21UE000XUS unit behind StorageReview’s benchmark results.
CPU performance and small-model inference
In StorageReview’s reported results, the P1 scored higher than the Dell Pro Max 14 Premium on Cinebench multi-core and the two listed inference tests, and completed the LLVM build in less time. These are results for the specific systems tested; the Dell’s RTX PRO 2000 ran at 47W, compared with 50W in the P1.
| Benchmark or workload | ThinkPad P1 Gen 9 (21UE000XUS, Core Ultra 9 386H, RTX PRO 2000 at 50W) | Dell Pro Max 14 Premium (Core Ultra 9 285H, RTX PRO 2000 at 47W) |
| Cinebench 2026 multi-core (points) | 5,083 | 4,292 |
| LLVM 21.1.4 full build (seconds) | 309.1 | 393.9 |
| Llama 3.1 8B, MLPerf Client (tokens/s) | 64.8 | 54.7 |
| Phi-4-mini, MLPerf Client (tokens/s) | 111.9 | 88.0 |
The P1 retained 4,811 Cinebench points after a roughly 30-minute loop, a reported 0.983 of its first-pass score. That sustained result adds useful context to the multi-core benchmark: the measured score stayed close to its initial run under this test’s conditions.
The GPU-memory boundary for local AI
StorageReview’s P1 ran the MLPerf Client v2.0 tests for Llama 3.1 8B and Phi-4-mini, but its larger model-load and coding-agent tiers did not run. The RTX PRO 2000 reported 7.7GB of GPU-usable memory; those test tiers began with gpt-oss:20b and required about 16GB.
The 64GB of system RAM and the GPU’s reported usable memory describe separate pools. For this tested configuration, the larger-tier constraint was the GPU-memory requirement in StorageReview’s suite. The result does not establish whether every model or context at a given size will fit; it describes the specific tests that ran and those that did not.
That makes the P1’s reported CPU performance relevant to multithreaded professional work, while local inference depends on the particular model-load requirement. Before choosing a configuration for local AI, compare its GPU-usable memory with the needs of the model you plan to run.
U.S. price observations
StorageReview reported a Lenovo.com list price of $7,719 on October 24, 2026, for a build matching the test system’s core configuration but with a Tandem OLED display. It reported a $5,399 sale price on October 22 for that OLED configuration. These dated prices were for the OLED build, not the reviewed IPS-panel unit.
StorageReview also listed a Core Ultra 7 356H configuration with an RTX PRO 2000, 32GB RAM, 1TB SSD, and IPS panel at $5,099; it reported that configuration’s sale price as $3,449 on October 22, 2026. Its reported $3,734 build-your-own starting price applied to a configuration with 16GB RAM, 512GB storage, and a WUXGA panel.
For buyers whose priority is CPU-heavy work or the smaller inference tests that ran, the reviewed configuration produced clear measured results. For local AI, the deciding specification is the GPU memory available to the target workload, not the system-RAM total alone.
