A laptop runs a block-building game alongside an AI companion, with a graphic explaining unified CPU and GPU memory.
Minecraft ran at around 120 FPS for much of the session, even with a 125-billion-parameter model running alongside it. That is XDA Developers' report from an HP OmniBook Ultra 16 with NVIDIA's RTX Spark. It is a single hands-on session, not a controlled benchmark. It does hint at what a big pool of unified memory could mean for Windows gaming and local AI on the same machine.

A laptop runs a block-building game alongside an AI companion, with a graphic explaining unified CPU and GPU memory. What was actually tested​

The HP OmniBook Ultra 16 in the test pairs a 20-core Arm CPU and a 6,144-core Blackwell GPU with 128GB of unified memory. The model was Qwen 3.8 Flash Next, a 125-billion-parameter mixture-of-experts model. The author used it to run a bot that joined a Minecraft world over LAN while he played.

The article's wording, "128GB of VRAM", needs a correction. RTX Spark uses a shared pool of unified memory. It isn't a discrete card's dedicated VRAM. StorageReview's lab test of the same laptop shows the difference. Its 128GB unit was configured as a 64GB GPU and 64GB CPU split, which measured 62.5GB of GPU-usable memory. The split is configurable, so the usable GPU share depends on how the machine is set up. XDA doesn't say how its unit was configured. A 125B model may also stream part of its data differently, as covered below.

How the bot works​

The bot runs on Mindcraft, an open-source project built on Mineflayer, a Node.js library for Minecraft bots. It joins the world like any other player and takes instructions through in-game chat. The model picks commands such as !collectBlocks("birch_log", 10). Mindcraft handles the movement, mining and inventory.

XDA describes the setup like this:

  1. Open the Minecraft world to LAN on a fixed port.
  2. Point Mindcraft at the world.
  3. Run node main.js, which also starts a web dashboard showing the bot's health, position and inventory.
  4. Use Mindcraft's LM Studio integration, which uses an OpenAI-compatible API with a configurable address, and aim it at a local llama.cpp server. Mindcraft has no dedicated llama.cpp option.

The author used NVIDIA's llama.cpp recipe for the model. It includes vision support and an MTP draft model for speculative decoding, and serves an OpenAI-compatible API on port 9931. Before launching Minecraft, he measured roughly 46 tokens per second at a short context, dropping to just under 42 at 32,000 tokens. Those are his own numbers, not official figures.

Why a 125B model isn't as heavy as it sounds​

"125B" is the total parameter count. It doesn't describe the compute used for each token. The Qwen3.8-Next architecture paper on arXiv describes a sparse mixture-of-experts model with 125B parameters and 6B activated per token. That is the reason a laptop can generate tokens at a usable speed. The paper also notes an extra 51B parameters of n-gram embedding tables held off the accelerator and prefetched from host memory. XDA says the same tables are streamed from storage rather than kept entirely in memory.

Other RTX Spark testing points the same way. MindStudio reported that mixture-of-experts models outperform dense ones on this hardware, with a 27B dense Qwen model reaching only 13.4 tokens per second. That was a different laptop, the MSI Prestige 16 AI Plus. Sparse models suit unified-memory machines better than dense ones of the same size.

Why Minecraft stayed fast​

The author names several reasons, and they come with caveats:

  • Native Arm64. He says Minecraft runs natively on Windows on Arm with Arm64 builds of Java and the graphics libraries, so it doesn't need x86 emulation through Microsoft's Prism.
  • No scheduling tricks. He says NVIDIA's CUDA in Graphics feature isn't currently supported on RTX Spark. So the llama.cpp server and the game shared the GPU without special optimization. I couldn't corroborate that claim in the other sources.
  • A light game. The author concedes that Minecraft isn't very demanding. He expects a title like Cyberpunk 2077 to suffer more with a 125B model running beside it. He has run that game at 1440p on a ThinkStation PGX using the same GB10 silicon.

The report leaves out the details that would turn "about 120 FPS" into a benchmark:

  • Minecraft edition and version
  • Resolution and render distance
  • Frame-time data
  • Model quantization
  • Power mode
  • A before-and-after FPS comparison

Treat it as a good sign, not a guarantee. Shared-memory contention shows up elsewhere in the data. StorageReview found that the OmniBook's browser slowed by 1.73x while a coding agent ran on the shared GPU. That was a different workload from Minecraft, but it shows that running both has a cost.

How well the bot played​

The model understood requests better than the bot carried them out:

  • Collecting logs and saplings: The bot gathered birch logs and dropped eight saplings at the player's feet.
  • Unique greetings: It generated a different greeting each time it joined or rejoined.
  • False promises: It said it would hand over logs it didn't have. The author had put it in Creative mode by mistake. It corrected itself after Mindcraft's logs showed it never received the items.
  • Disabled code execution: It repeatedly tried to call !newAction even though the author had disabled it. It then fell back to standard commands.
  • Falls and getting stuck: It jumped into a hole and disconnected, then rejoined. While digging dirt it announced it was stuck and disconnected again. Mindcraft's unstuck system detects 20 seconds without movement and tries to move five blocks away. If that fails after another ten seconds, it terminates the agent and retries.

The author blames these problems on Mineflayer pathfinding and Mindcraft's execution layer. He says they aren't down to the model or to performance. That is plausible, but the report shows no testing that separates the causes.

A security note​

!newAction lets the model run JavaScript on your computer. The author left it off, and he is right to. An LLM that can run code on a Windows machine needs a sandbox and careful permissions, which isn't what you want for a Minecraft experiment. NVIDIA's RTX Spark launch materials mention secure runtimes and execution containers to control agent permissions. That makes the same point: local agents need boundaries.

What this does and doesn't show​

What it shows:

  • A 125B-total-parameter sparse model can run locally on a Windows on Arm laptop at roughly 40 to 46 tokens per second in the author's test.
  • A light game can stay near 120 FPS in the same session.
  • No cloud API was involved.

What it doesn't show:

  • That demanding games will hold their frame rates.
  • That other RTX Spark memory configurations will behave the same way.
  • That this is a good way to play Minecraft with an AI companion.

Price matters too. StorageReview says HP listed the 128GB, 2TB build at $6,229.99 as of October 14. That is a lot to pay for a Minecraft companion. The 32GB model cannot load StorageReview's model tiers. Treat the 125B result as belonging to the top configuration.

Bottom line​

Unified memory changes the usual local-AI-versus-gaming tradeoff. On a 16GB GPU, the model and the game fight over memory. Here the question is how the 128GB pool is split and how much bandwidth the two workloads share. The XDA experiment is informal, but it gives a concrete data point. A sparse 125B model and a light game can coexist on one Windows laptop. Heavier games and other model and memory combinations remain untested.

 

References

  1. I played Minecraft with a 125B local LLM on an RTX Spark laptop, and the game still held 120 FPS XDA 2026-10-25T11:00:20+00:00
  2. HP OmniBook Ultra 16 Review: RTX Spark With 128GB Tops Our Local Coding Agent Test - StorageReview.com storagereview.com
  3. Running 120B+ LLMs Locally on the RTX Spark Laptop: Real Numbers mindstudio.ai