Project Maya enables local execution of DeepSeek-V4.1-Flash and GLM-5.3-Flash models on consumer hardware

Project Maya version 1.1.16 has introduced support for the DeepSeek-V4.1-Flash (1017 checkpoint) and GLM-5.3-Flash models, enabling local inference on consumer-grade hardware. In testing, the 748B parameter DeepSeek-V4.1-Flash model achieved approximately 11.1 tokens per second on a Dell Precision T5810 workstation equipped with dual NVIDIA RTX 3060 GPUs and 256GB of RAM. The GLM-5.3-Flash model demonstrated higher performance, reaching speeds between 46 and 60 tokens per second on a more powerful multi-GPU rig. The software, a fork of Strata and Llama.cpp, utilizes model-specialized runtimes and custom quantization, such as the Maya-S version for DeepSeek. These optimizations allow for significant pre-fill speeds of 1,100 to 1,500 tokens per second within a 1-million-token context window. The creator showcased the practical utility of these models by developing "Turret Control: Operation Lumberjack," a World War II-themed on-rails shooter game, using GLM-5.3-Flash. This development process involved roughly 350 million tokens and 56 code commits over several days. Comparative testing of SVG generation tasks showed that GLM-5.3-Flash produced more detailed animations than DeepSeek-V4.1-Flash, though both models successfully rendered complex prompts locally.

Tech4K Staff8:31 watch

Project Maya version 1.1.16 has introduced support for the DeepSeek-V4.1-Flash (1017 checkpoint) and GLM-5.3-Flash models, enabling local inference on consumer-grade hardware. In testing, the 748B parameter DeepSeek-V4.1-Flash model achieved approximately 11.1 tokens per second on a Dell Precision T5810 workstation equipped with dual NVIDIA RTX 3060 GPUs and 256GB of RAM. The GLM-5.3-Flash model demonstrated higher performance, reaching speeds between 46 and 60 tokens per second on a more powerful multi-GPU rig. The software, a fork of Strata and Llama.cpp, utilizes model-specialized runtimes and custom quantization, such as the Maya-S version for DeepSeek. These optimizations allow for significant pre-fill speeds of 1,100 to 1,500 tokens per second within a 1-million-token context window. The creator showcased the practical utility of these models by developing "Turret Control: Operation Lumberjack," a World War II-themed on-rails shooter game, using GLM-5.3-Flash. This development process involved roughly 350 million tokens and 56 code commits over several days. Comparative testing of SVG generation tasks showed that GLM-5.3-Flash produced more detailed animations than DeepSeek-V4.1-Flash, though both models successfully rendered complex prompts locally.

Project Maya v1.1.16 adds local support for DeepSeek-V4.1-Flash and GLM-5.3-Flash models. DeepSeek-V4.1-Flash runs at approximately 11 tokens per second on dual NVIDIA RTX 3060 GPUs.

GLM-5.3-Flash achieves inference speeds between 46 and 60 tokens per second on high-end consumer rigs. The software supports a 1-million-token context window for complex coding and development tasks.

Pre-fill performance for large context windows ranges from 1,100 to 1,500 tokens per second. The creator utilized GLM-5.3-Flash to generate a functional 3D game through 56 code commits.

Comparative testing indicates GLM-5.3-Flash currently produces superior SVG animations compared to DeepSeek-V4.1-Flash.

Chapter guide

Introduction to Project MayaOverview of the latest Project Maya release and its support for new Flash models.Software Lineage and New ModelsExplanation of Maya as a fork of Strata and its implementation of specialized runtimes.Unsloth and Llama.cpp AlternativesBrief comparison with other local LLM runners and quantization methods like UDIQK1M.GLM-5.3-Flash Performance TestLive demonstration of GLM-5.3-Flash inference speeds and token generation.Game Development ShowcaseDemonstration of a game built entirely using local AI models and 350 million tokens.Pre-fill and Context WindowTechnical details on pre-fill speeds and the 1-million-token context window capability.SVG Generation ComparisonSide-by-side comparison of GLM-5.3-Flash and DeepSeek-V4.1-Flash generating animated SVGs.Conclusion and Hardware RigsFinal thoughts on local AI accessibility and references to different hardware configurations.

Worth noting

  • Performance metrics are based on specific hardware configurations, including a Dell Precision T5810 with dual RTX 3060s and a separate high-end multi-GPU rig.
  • The DeepSeek-V4.1-Flash model demonstrated a very long 'thinking' time of over 4,400 seconds for a complex prompt on the lower-end hardware.
  • The game 'Turret Control' was developed by the video creator, and its performance may vary based on the browser and local hardware.

Watch the original video ↗