Older Video Game Consoles Can Now Run Modern Artificial Intelligence Offline
San Francisco, Tuesday, 17 November 2026.
Software developers have successfully run modern nine-billion-parameter artificial intelligence on legacy 2013 game consoles, proving that low-cost, repurposed hardware can deliver high-speed computing without internet access or expensive infrastructure.
The Garlic LLM Breakthrough: Running 9B AI Models on 2013 Hardware
On November 15, 2026, open-source developer Leonardo Alegría released Garlic LLM [1][2][5], an open-source framework enabling legacy PlayStation 4 consoles to execute modern nine-billion-parameter AI models such as Alibaba’s Qwen3.5-9B entirely offline [1][2][5]. The breakthrough demonstrates that 2013-era hardware, originally designed for gaming, can now perform high-throughput edge inference at speeds up to 43 tokens per second [1][2][5]. By bypassing traditional operating system drivers and communicating directly with the console’s AMD Radeon-based GPU through custom PM4 command buffers, Garlic LLM achieves performance levels previously thought impossible on consumer hardware from that generation [1][2][5]. This marks a significant milestone in bare-metal programming and hardware optimization, offering a glimpse into cost-efficient edge AI deployment strategies for enterprises and hardware manufacturers alike.
Technical Ingenuity: How a Jailbroken PS4 Handles a 9B Model
The technical feat hinges on three core innovations: aggressive model quantization, direct GPU access, and speculative decoding [1][2][5]. The Qwen3.5-9B model, which would typically require ~18 GB of memory in FP16 precision, is compressed to approximately 4.5 GiB using mixed 3-bit and 4-bit quantization, fitting within the PS4’s 8 GB shared GDDR5 memory pool [1][2][5]. Leonardo Alegría’s software bypasses Sony’s graphics drivers entirely, instead writing low-level PM4 command buffers by hand to interface directly with the GPU’s Graphics Core Next (GCN) architecture [1][2][5]. This bare-metal approach contrasts sharply with conventional inference frameworks like llama.cpp, which rely on higher-level APIs that are unavailable on the PS4’s homebrew ecosystem [1][2][5]. Additionally, the implementation leverages Qwen3.5’s multi-token prediction—speculative decoding—to increase inference throughput from ~22 tokens per second to 31–39 tokens per second [1][2][5].
Performance Benchmarks and Feature Set
Under real-world testing, Garlic LLM achieves 31 to 39 tokens per second during regular text chat and up to 43 tokens per second during code generation on a base or Slim PS4 running firmware 11.52 [1][2][5]. Experimental results on a PS4 Pro indicate potential speeds up to 56 tokens per second in native mode, though this configuration remains untested and unsupported due to architectural differences [1][2][5]. The application supports voice interaction through integrated models like Parakeet (0.33 GiB) for dictation and Chatterbox (0.29 GiB and 0.09 GiB) for speech synthesis in seven languages, alongside MicroPython code execution for local scripting [1][2][5]. Despite these capabilities, enabling voice features reduces generation speed to approximately 22 tokens per second due to additional CPU load [1][2][5]. All processing occurs offline, with no internet connection required, and the entire 5.66 GB package can be installed via the PS4’s Package Installer with just 6 GB of free space, or up to 12 GB if pre-copied to internal storage [1][2][5].
Implications for Edge AI and Legacy Hardware Markets
The Garlic LLM project represents more than a technical curiosity—it signals a paradigm shift in edge AI deployment, particularly in sectors where cost, durability, and offline operation are critical [1][2][3][5]. For enterprises evaluating edge inference infrastructure, the ability to repurpose aging consumer hardware reduces capital expenditure while extending the usable lifespan of silicon assets by up to a decade [1][2][3][5]. This trend aligns with growing demand for sovereign AI systems in government, defense, and industrial applications, where data sovereignty and operational continuity outweigh raw performance metrics [1][2][3][5]. Moreover, the project underscores the strategic value of open-source collaboration in hardware optimization, with the developer actively soliciting user feedback via GitHub to refine performance across diverse PS4 variants [1][2][5]. While the PS4 Pro remains untested and the PS5 is intentionally unsupported due to architectural incompatibilities, the initiative opens new avenues for retrofitting legacy devices across consumer electronics, automotive systems, and embedded platforms [1][2][5]. In a broader context, this innovation challenges the assumption that AI acceleration requires cutting-edge GPUs or cloud connectivity, instead proving that software ingenuity can unlock latent capacity in existing hardware inventories [1][2][3][5].
Sources
- www.tomshardware.com
- mezha.net
- www.tweaktown.com
- www.facebook.com
- hwbusters.com
- x.com
- huggingface.co
- x.com