Latest in AI
Maya-XS Runs DeepSeek V4.1 Flash in 120.6 GB: What Still Lives on SSD
Project Maya's new experimental quantization fits every expert into a 128 GB Ryzen AI Max, but the memory headline needs a storage footnote.
Project Maya released Maya-XS on October 22, an experimental quantization of DeepSeek V4.1 Flash that the maintainer says uses 120.6 GB of memory, including its drafter. On a 128 GB Ryzen AI Max, every expert can stay in GPU memory. That does not mean the entire system fits there.
What Maya-XS actually changes
The v1.1.22 release notes describe roughly 1.6-bit IQ1_S experts, with all 384 experts per layer retained. The separate Engram tables remain on SSD: another 161 GB in Q6_K format. Maya-XS includes a two-bit DSpark drafter.
The update also adds DeepSeek tensor parallelism across multiple GPUs and improves single-GPU drafting. Those performance claims come from specific maintainer test setups, not a guarantee for every PC.
Expert residency and complete model residency are different things. The useful change is reducing how much expert data must move through slower storage, while retaining a separate disk-backed component. Calling this a 120.6 GB download would obscure that distinction.
How this differs from ordinary offloading
The Project Maya repository describes a tiered engine: frequently used experts stay on GPU, others sit in RAM, and the remainder are read from NVMe storage. It exposes OpenAI- and Anthropic-compatible APIs with tool calls, so local inference can sit behind existing coding-agent workflows.
Maya-XS changes the capacity calculation for one particular DeepSeek configuration. It should not be read as a general claim that any 128 GB machine delivers the same experience. Memory available to the GPU, the operating system, the active context and disk performance still matter.
The repository also distinguishes its model variants. Its published next-token agreement figure for DeepSeek Maya-S is not a quality score for Maya-XS. A smaller quantization needs its own evidence before those results can be carried across.
What to check before relying on it
For a developer evaluating local coding, the useful test is whether the smaller model completes the same tasks, not whether a short answer looks fluent. Run representative tool calls and longer debugging sessions against a higher-precision baseline where possible. Keep latency, memory use and task success separate.
The AMD setup documentation labels the platform path experimental and lists specific supported GPU families and software requirements. That makes this an evaluation option rather than a reason to buy hardware from the headline alone.
The broader opportunity is practical: fitting the experts into an existing machine can make private experimentation easier. The next meaningful result would be a clear quality-and-speed comparison for Maya-XS itself, with context length and SSD configuration included. Capacity opens the door; reliable task completion decides whether it stays useful.
I'm Paras. I work at the intersection of AI, engineering and distribution.
I help technical founders turn model releases and engineering updates into clear stories that developers can understand and share.