Artificial Analysis published its first AA-AgentPerf-Local results for an NVIDIA RTX Spark N1X laptop on October 15, 2026. The laptop completed the recorded agent workload on Qwen3.8 27B in 18.7 minutes. The RTX Spark N1X combines a Blackwell GPU with up to 128 GB of shared memory in a Windows laptop.

AA-AgentPerf-Local replays 8 recorded agent tasks—5 from PinchBench and 3 from SWE-bench—across 168 model turns, each carrying the conversation so far. The longest task reaches 56K context tokens. The workload sends 2.6M prompt tokens, including 196K new prompt tokens while the rest repeat earlier content, and 31K output tokens; each system generates the same number of output tokens. Most prompts repeat previous content, and a server that retains its cache reads the opening prompt and a median of about 450 new tokens per turn, rising to 6.4K after a large tool output.

The leaderboard tests three models at 4-bit precision on five systems: the RTX Spark laptop, NVIDIA DGX Spark, AMD Ryzen AI Halo, MacBook Pro with M5 Pro, and a GeForce RTX 5090 desktop. The RTX Spark N1X and DGX Spark share a 20-core Arm CPU and 6,144-core Blackwell GPU; the laptop has 128 GB LPDDR5X memory at 300 GB/s, versus 273 GB/s on DGX Spark. NVIDIA gives the N1X laptop chip a 45–80 W power range, while GB10 in DGX Spark is rated at 140 W. Artificial Analysis tested a prototype 15-inch Surface Laptop Ultra with 128 GB, priced at $5,899.99, running Windows 11 on Arm and a native Windows build of llama.cpp on the GPU. It was plugged in and set to high performance.

On Qwen3.8 27B, the RTX Spark finished in 18.7 minutes, near DGX Spark at 17.3 minutes, and ahead of MacBook Pro at 37.6 minutes and Ryzen AI Halo at 34.5 minutes; the RTX 5090 desktop led at 4.9 minutes. On Qwen3.6 35B A3B, which activates 3B parameters per token, it took 7.6 minutes, compared with 7.3 for DGX Spark, 11.7 for Ryzen AI Halo, and 11.9 for MacBook Pro. On Ling 3.0 Flash, it took 15.3 minutes versus 14.9 for DGX Spark and 25.1 for Ryzen AI Halo; its 77 GB of weights do not fit on the MacBook Pro or RTX 5090.

Each system uses its own runtime, speculative decoding method, and 4-bit weight format. The RTX Spark uses NVIDIA’s llama.cpp recipe with UD-Q4_K_M weights and DFlash with 7 draft tokens; DGX Spark uses the same. The MacBook Pro uses Q4_K_M and MTP with 4 draft tokens, while Ryzen AI Halo and RTX 5090 use Q4_K_M and MTP with 2 draft tokens. Artificial Analysis is testing smaller formats IQ4_XS and Inco AI’s Splash, a Metal engine for Apple Silicon using DFlash, and says it will publish results after accuracy checks.

At approximate October 2026 prices, the RTX Spark laptop at $5,899.99 costs more than the MacBook Pro at about $3,700 and Ryzen AI Halo at $4,000, but slightly less than DGX Spark at about $6,000. The RTX 5090 desktop is estimated at about $9,500, including a roughly $7,000 graphics card and an estimated $2,500 for the rest of the PC. The leaderboard allows users to enter their own system prices.

Artificial Analysis published its first AA-AgentPerf-Local results for an NVIDIA RTX Spark N1X laptop on October 15, 2026. The laptop completed the recorded agent workload on Qwen3.8 27B in 18.7 minutes. The RTX Spark N1X combines a Blackwell GPU with up to 128 GB of shared memory in a Windows laptop.

AA-AgentPerf-Local replays 8 recorded agent tasks—5 from PinchBench and 3 from SWE-bench—across 168 model turns, each carrying the conversation so far. The longest task reaches 56K context tokens. The workload sends 2.6M prompt tokens, including 196K new prompt tokens while the rest repeat earlier content, and 31K output tokens; each system generates the same number of output tokens. Most prompts repeat previous content, and a server that retains its cache reads the opening prompt and a median of about 450 new tokens per turn, rising to 6.4K after a large tool output.

The leaderboard tests three models at 4-bit precision on five systems: the RTX Spark laptop, NVIDIA DGX Spark, AMD Ryzen AI Halo, MacBook Pro with M5 Pro, and a GeForce RTX 5090 desktop. The RTX Spark N1X and DGX Spark share a 20-core Arm CPU and 6,144-core Blackwell GPU; the laptop has 128 GB LPDDR5X memory at 300 GB/s, versus 273 GB/s on DGX Spark. NVIDIA gives the N1X laptop chip a 45–80 W power range, while GB10 in DGX Spark is rated at 140 W. Artificial Analysis tested a prototype 15-inch Surface Laptop Ultra with 128 GB, priced at $5,899.99, running Windows 11 on Arm and a native Windows build of llama.cpp on the GPU. It was plugged in and set to high performance.

On Qwen3.8 27B, the RTX Spark finished in 18.7 minutes, near DGX Spark at 17.3 minutes, and ahead of MacBook Pro at 37.6 minutes and Ryzen AI Halo at 34.5 minutes; the RTX 5090 desktop led at 4.9 minutes. On Qwen3.6 35B A3B, which activates 3B parameters per token, it took 7.6 minutes, compared with 7.3 for DGX Spark, 11.7 for Ryzen AI Halo, and 11.9 for MacBook Pro. On Ling 3.0 Flash, it took 15.3 minutes versus 14.9 for DGX Spark and 25.1 for Ryzen AI Halo; its 77 GB of weights do not fit on the MacBook Pro or RTX 5090.

Each system uses its own runtime, speculative decoding method, and 4-bit weight format. The RTX Spark uses NVIDIA’s llama.cpp recipe with UD-Q4_K_M weights and DFlash with 7 draft tokens; DGX Spark uses the same. The MacBook Pro uses Q4_K_M and MTP with 4 draft tokens, while Ryzen AI Halo and RTX 5090 use Q4_K_M and MTP with 2 draft tokens. Artificial Analysis is testing smaller formats IQ4_XS and Inco AI’s Splash, a Metal engine for Apple Silicon using DFlash, and says it will publish results after accuracy checks.

At approximate October 2026 prices, the RTX Spark laptop at $5,899.99 costs more than the MacBook Pro at about $3,700 and Ryzen AI Halo at $4,000, but slightly less than DGX Spark at about $6,000. The RTX 5090 desktop is estimated at about $9,500, including a roughly $7,000 graphics card and an estimated $2,500 for the rest of the PC. The leaderboard allows users to enter their own system prices.