In an account published Nov. 6, Ed Bott says an HP OmniBook Ultra with NVIDIA RTX Spark handled a local coding-and-archiving project across 479 inference calls and nearly 50 million logical tokens, with $0 in inference charges. That figure covers inference for his project, not the laptop or its electricity use.
Treat a $0 inference bill as one cost figure, not the total cost of the system.
A coding agent built Bott’s archive
Bott says Microsoft Copilot helped him set up the open-source inference runtime llama.cpp, the open-weights model Qwen3.8-27B, and Cline inside Visual Studio Code. He supplied a detailed application specification, then let Cline write code, run tests and revise its work, with occasional human approval. Bott says the setup took less than half an hour.
The task started with a spreadsheet of article links Bott had assembled over two years. The project retrieved nearly all the listed articles. For roughly 90 links that returned 404 errors, it tried recovery through the Internet Archive’s Wayback Machine; roughly 90 more pages remained unretrieved. Bott says he spent about 30 minutes reviewing and resolving the remaining problems.
The resulting archive contained 1,951 raw-HTML files and 1,951 Markdown files, plus a script for adding future posts. Bott’s account gives a concrete example of a coding agent handling a bounded, specified task through repeated code-and-test cycles. Its token and call totals describe this project, rather than a standardized performance comparison.
Bott also reported audible fan noise while the model was working. He said the laptop did not become uncomfortably hot and found a desk more suitable than his lap for this workload.
What the $0 figure covers, and which HP configuration is listed
Bott described his test unit as having 128 GB of unified memory. HP’s US store lists a separate configuration, SKU E5AW2UA#ABA, with 64 GB of RAM and a 2 TB SSD. On Nov. 8, HP listed that SKU for $4,399.99. The displayed price belongs to the 64-GB configuration, not Bott’s described test unit. HP’s US product listing
The $0 result refers to inference charges for the project. It does not cover the computer’s purchase price or provide an electricity-inclusive total.
For a separate look at local-model access on this platform, see our earlier HP RTX Spark and Hermes report. Bott’s project focused on building an article archive with a coding agent.
Why local AI users value it differently
In a Reddit discussion about local AI, some participants described privacy, control, learning and hobby enjoyment as reasons to run models locally. Others cautioned against buying new hardware solely for financial savings. Those comments concerned varied setups, not Bott’s particular HP laptop.
