A personal demo has surfaced showing someone connecting DeepSeek's new model to a tool-executing AI agent to shrink down Windows' "BIOHAZARD 7 resident evil" for Android devices. The poster explains they reduced the texture files from 20GB to 8GB, bringing the entire game down to 13GB. Wccftech followed up with additional reporting, stating that performance on a OnePlus 12R improved from roughly 20FPS to a stable 30FPS. However, this was not a pure AI performance test. Texture reduction and SGSR were used together, and conditions like internal resolution and thermal state before and after the change remain unknown—so the results deserve careful interpretation.
What Did the AI Agent Actually Change?
The story began with an X post on September 14, 2026. The poster explicitly states they used DeepSeek V4.1 Flash and Hermes Agent, and that they compressed 20GB of texture files down to 8GB. They note the modified game totals 13GB, while the iOS version is approximately 25GB. That's as far as the original post goes—no prompts, conversion scripts, change logs, or execution records were shared.
The following day, September 15, Wccftech reported additional details obtained directly from the poster. They connected the OnePlus 12R to a Mac via USB for debugging, reduced 4,096-pixel textures down to 1,024 or 512 pixels, and removed some audio and skin files. Using Snapdragon Game Super Resolution (SGSR) to upscale from 720p to 1200p, performance reportedly went from a barely-there 20FPS to a stable 30FPS. Power consumption was reported at 5-6W. None of these figures appear in the original X post, so they should be treated as the poster's explanation obtained through Wccftech's follow-up reporting, not values verified through publicly available logs.
It's worth separating out what role the AI actually played. DeepSeek V4.1 Flash is a Mixture-of-Experts model released on September 10, 2026. Hermes Agent is not a model itself—it's an execution framework that gives an LLM tools for device operations, file manipulation, and code execution. In this setup, DeepSeek appears to have handled diagnosis and procedure generation, while Hermes mediated the external operations. However, which tools were enabled and where a human approved or corrected the process remains undisclosed. This cannot be confirmed as fully autonomous optimization.
Where Does the Load Accumulate in Winlator?
The OnePlus 12R runs a Snapdragon 8 Gen 2, Adreno 740 GPU, and either 8GB or 16GB of LPDDR5X memory. Which memory configuration was used, along with the Android and GPU driver versions and available storage, are all unknown. The device was announced on January 23, 2024, meaning roughly 2 years and 8 months had passed by the time of this demo. Wccftech's headline describes it as a "four-year-old smartphone," which doesn't match the device's actual age. This appears closer to a rounded reference to the Snapdragon 8 Gen 2's chip generation, which was announced on November 15, 2022—about 3 years and 10 months prior.
Running a PC title on this device involves a long chain of translation layers. Winlator combines Wine with Box86/Box64 to run Windows x86_64 applications on Android. Box64 dynamically translates x86_64 instructions on ARM64 hardware, supporting Wine64. If the game uses Direct3D 8-11, DXVK translates the rendering commands into Vulkan, which the Adreno driver then executes on the GPU. This creates a chain: CPU instruction translation, Windows API compatibility, graphics API translation, asset loading, and GPU processing.
Given this, the I/O or memory bandwidth constraints that Wccftech attributed to the poster's explanation are technically plausible. However, no profiler results have been made public. Shrinking textures can reduce not just storage but also load times and GPU memory transfer volume. On the other hand, removing audio and skins means the content is no longer identical to the original—this is not an apples-to-apples comparison preserving the official version's quality and content. Steam's current minimum requirements for the PC version call for 24GB of free space and 1080p/30FPS, so the poster's 13GB figure doesn't imply the official distribution's requirements have changed.
Why the "50% Improvement" Can't Be Definitively Attributed to AI
The increase from 20FPS to 30FPS represents a 50% gain: (30-20)÷20=50%. But it's unclear whether the same scene was run for the same duration, whether Winlator, Wine, Box64, DXVK, and GPU driver versions matched before and after, or whether internal resolution and device temperature were controlled for. No average FPS, 1% low values, frame time data, or run counts have been published. While DXVK does have features to display FPS and frame time, no raw logs from this demo exist.
The original post's texture figures—20GB down to 8GB—represent a 60% reduction. The 68% figure in Wccftech's headline is calculated from 25GB to 8GB, but the 25GB figure was originally described in the source post as the iOS version's size. The math works out to (20-8)÷20=60% and (25-8)÷25=68%, respectively. However, both figures are self-reported by the poster, with no file listings or actual byte counts released. The iOS version and the modified Windows version also differ in distribution format and bundled content, so comparing 25GB against 8GB doesn't yield an apples-to-apples compression ratio.
The 5-6W figure also lacks a clear denominator. It's not specified whether this measures the SoC alone, the entire device, a software estimate, or an external measurement device. Without a pre-modification power figure, there's no way to calculate what percentage efficiency improved. When multiple changes are applied simultaneously, it becomes impossible to isolate how many FPS came from DeepSeek's decisions, texture reduction, file deletion, SGSR, or reduced thermal throttling.
What "1200p via SGSR" Doesn't Mean
SGSR1 is a single-pass spatial upscaler that uses the current frame as input. It combines upscaling (the process of expanding a lower internal resolution to a higher output resolution) and sharpening into a single shader pass, converting an image rendered at lower resolution into a larger output. This differs from SGSR2, which uses temporal information.
So "720p to 1200p" doesn't mean every pixel of the 1200p image was rendered from scratch. It should be read as: the GPU load was reduced by rendering at a lower internal resolution, then upscaled to roughly 1200p output. While Qualcomm touts performance gains for SGSR generally, those demo figures can't be directly transferred to this particular Windows-game execution pipeline, since which version and shader settings were integrated into the Winlator side remain undisclosed.
Generally speaking, reducing texture size and lowering internal resolution affect different types of load. The former can impact asset size, loading, and memory usage. The latter primarily reduces the number of pixels being rendered. With texture reduction and SGSR both applied simultaneously as in this case, it's impossible to isolate the primary cause of reaching 30FPS or determine the visual quality impact. Beyond static image resolution, checking for flickering or edge stability during motion would also be necessary.
What It Would Take to Make This a Reproducible Result
The value of this case lies in demonstrating that a general-purpose AI model paired with a tool-execution framework can potentially link together game asset investigation, conversion work, and on-device verification. But turning this into a reproducible technical achievement requires more than a successful demo video. The prompts and tools used, conversion scripts, target file lists, versions spanning from Winlator to GPU drivers, game settings, and internal/output resolutions all need to be disclosed together.
Measurement also needs to be standardized. Running multiple trials with the same save data and scene, the same movement path, the same starting temperature, and the same measurement duration, while recording average FPS, 1% low values, frame time distribution, temperature, clock speeds, and total device power consumption for each change. Adding one variable at a time—first textures alone, then file deletion, then SGSR—would allow tracking which change improved what.
This demo shows real potential for AI agents to assist with complex real-device optimization. But turning that potential into a validated product assessment requires more than a model name or a single 50% gain—it requires third parties being able to reproduce the same results from the same inputs and procedures.