Key Findings
Project Maya v1.1.28, released on October 26, 2026, significantly accelerates text generation for DeepSeek models in Linux environments using a single NVIDIA GPU. In tests using an RTX 3090, generation speed increased from 12.9 to 27.8 tokens/sec. However, this improvement primarily affects the generation phase; prefill speeds remained nearly unchanged, with 256K token inputs processing at 1,407 tokens/sec compared to 1,417 tokens/sec in the previous version.
Technical Details
The speedup is achieved by optimizing the data supply path from RAM to GPU, allowing the GPU to reference RAM data directly and bypassing intermediate buffers. This reduces GPU idle time during DSpark speculative decoding by caching loaded experts in VRAM. The update also introduces support for 512K and 1M context lengths. Testing with nine GPUs and a 3.5bit quantized model confirmed the ability to retrieve embedded codes from 1,028,310 tokens, though this required 23.8 minutes for processing at a speed of 720 tokens/sec.
Background & Context
Project Maya is designed to run Mixture of Experts (MoE) models—such as the 748 billion parameter DeepSeek-V4.1-Flash-1017 and the 321 billion parameter GLM-5.3-Flash—by distributing experts across VRAM, system RAM, and NVMe SSDs. For example, Maya-S for DeepSeek requires approximately 340GB of memory. The 1M context support increases VRAM consumption by approximately 1.1KB per token per GPU in a tensor parallel group, which may compete with expert caches and reduce overall generation speed.
Significance & Outlook
Performance gains are hardware-dependent. While some users saw improvements, another reported a 19% performance drop on an NVIDIA L40S (48GB) in a virtual machine environment, where CPU direct calculation proved faster than the new GPU transfer path. Additional updates include RDNA-optimized kernels for AMD GPUs and distributed speculative decoding for multi-GPU DeepSeek configurations. Users are advised to monitor startup logs to verify if the new data path is active and to measure prefill versus generation speeds independently.
Source: https://xenospectrum.com/project-maya-deepseek-decode-context/
Get our weekly technology intelligence — free
Receive an infographic that lets you judge at a glance whether each field’s analysis report is worth reading.
Subscribe Free — Weekly Tech Intelligence
By subscribing, you’ll receive Troy-Technical’s weekly technology intelligence newsletter.
- Your email and selected fields are used only to deliver the newsletter.
- We never share your information with third parties.
- You can unsubscribe anytime via the link in each email.
See our Privacy Policy for details.
Takes about a minute · Unsubscribe anytime
