
China's most watched domestic GPU maker is coming back to the consumer market, and its pitch is aimed squarely at the one number every local-AI enthusiast cares about: VRAM. Moore Threads has teased a new desktop graphics card codenamed Lushan, which it says will carry the largest video memory capacity of any consumer GPU ever made. The full reveal is scheduled for December 29 at the company's MUSA Developer Conference in Beijing — but the exact number is still a secret, and the gap between what is confirmed and what is rumored is worth understanding before anyone gets excited.
The short version: Lushan is real, the record-VRAM claim comes from the company itself, and the 48GB and 64GB figures circulating are educated guesses, not specifications. What makes the story interesting anyway is why a Chinese GPU maker would lead with memory instead of raw speed — and what that says about who this card is really for.
The teaser surfaced on November 18, when a product manager for Moore Threads' MTT S80 posted on social media that the company's next consumer card would have the highest VRAM capacity in consumer graphics card history. The company followed up with an official poster for MDC 2026, its MUSA Developer Conference, scheduled for December 29–30 in Beijing. The artwork shows two mountain peaks — one a 3D wireframe mesh, the other styled as a neural network — with the card itself mostly hidden behind them.
Coverage from Wccftech, Hardware Busters, and GameGPU agrees on the basics: the Lushan name, the record-VRAM claim, a December 29 reveal, and a positioning that covers both AAA gaming and local AI inference. The teaser also mentions a new rendering architecture called UNITE, a newly developed ray tracing engine, support for AI-generated rendering, and performance claims of 16 times the geometry processing, 50 times the ray tracing performance, and 64 times the AI computing capability of the current MTT S80 and S90 cards.
Notice what is missing from that list: capacity, memory type, bus width, bandwidth, power draw, price, and benchmarks. All of it is being held back for the conference. So where do the numbers in the headlines come from?
Two different numbers are floating around, and they come from two different kinds of guesswork.
The 48GB figure traces back to reporting that the card would ship with "48GB or more" of VRAM, putting it directly against Nvidia's flagship RTX 5090, which carries 32GB of GDDR7. That figure has been repeated widely, but no official specification backs it yet — treat it as a rumor with a plausible source, not a fact.
The 64GB figure is a different kind of speculation. Moore Threads previously said its next generation would carry four times the memory of the MTT S80, which ships with 16GB of GDDR6. Four times 16 is 64. If the company follows through on that earlier statement, 64GB is the arithmetic — but it is still an inference from an old roadmap promise, not a published spec.
Either number would be a record for a consumer card. The honest summary is: Moore Threads claims the record, the industry expects something above 32GB, and nobody outside the company knows the exact figure. December 29 settles it.
For gamers, more VRAM mostly means higher resolutions and bigger texture packs. For the local-AI crowd — the audience Moore Threads is clearly courting — VRAM is the hard ceiling on what you can run at all.
A large language model has to live in GPU memory while it runs. A 27-billion-parameter model quantized to 8 bits needs roughly 27–30GB of framebuffer by one estimate, which means it barely fits on an RTX 5090 and not at all on a 24GB RTX 4090. Bump the model to 32 billion parameters, or run it at higher precision, and even the 5090 is out. That is why the local-model community on forums like LINUX DO reacted to the Lushan teaser with a very specific kind of interest: not "will it play games well," but "will it let me run a bigger model on one card, at home, without renting cloud GPUs?"
This is also why the comparison that matters is not against the RTX 5090's raw speed. Nobody seriously expects Lushan to out-benchmark Nvidia's flagship on CUDA workloads or ray tracing. The comparison that matters is memory per dollar: if a consumer card offers 48GB or more at a price anywhere near a high-end gaming GPU, it becomes the cheapest way to run models that today require a data-center card or a multi-GPU rig.
To judge the teaser, it helps to know where Moore Threads started. The MTT S80, launched in November 2022, was China's first proper gaming GPU: 4,096 MUSA cores at a 1.8GHz boost clock, 14.7 TFLOPS of FP32 compute, 16GB of GDDR6 on a 256-bit bus with 448 GB/s of bandwidth, and a 255W power draw, all on a PCIe 5.0 interface. On paper the specs looked respectable. In practice, independent testing found the card competitive in esports titles at high resolutions but trailing an RTX 3060 in demanding games — and reviewers consistently pointed at the same culprit: immature drivers, not silicon.
That driver story is the shadow hanging over Lushan. A GPU is only as good as its software stack, and Moore Threads' MUSA platform — its CUDA alternative — has spent four years catching up. The company has been building the full stack in parallel: the MUSA SDK for developers, the KUAE computing cluster platform for AI training, and a growing list of domestic data-center deployments. Lushan is the first consumer card to benefit from all of that work, and the teaser leans on it: the UNITE architecture, the new ray tracing engine, and AI-generated rendering are all software stories as much as hardware ones.
The context is geopolitical as well as technical. Moore Threads was added to the U.S. Entity List in October 2023, which restricted its access to leading-edge chip manufacturing using U.S. technology. The company has continued to ship products anyway, leaning on domestic demand and older process nodes. Lushan is, in that sense, a statement product: a Chinese consumer GPU built under sanctions, aimed at the exact workload — local AI — where export controls on Nvidia hardware have created the most domestic demand.
If you are following this story, here is what to actually watch for when the reveal happens:
Lushan is a useful lens on a shift that is bigger than any single product. Nvidia and AMD have spent the last two years separating their gaming cards from their AI accelerators, keeping the biggest memory pools on data-center parts that cost many times more. That leaves a gap: a large and growing group of developers and enthusiasts who want to run serious models locally, on hardware they own, and cannot justify data-center prices.
Chinese GPU makers see that gap clearly, because export controls made it their home market's defining problem. Moore Threads is not the only company chasing it — domestic rivals and open-source efforts like Tilelang are building the software and hardware layers around the same idea — but Lushan is the most concrete consumer product aimed at it so far. If the card delivers on even half of its memory promise, with usable drivers, it changes the math for local AI in China and gives the rest of the world a reason to pay attention to a GPU maker most people outside China have never heard of.
Until December 29, though, the right attitude is the one the LINUX DO discussion itself landed on: interested, skeptical, and waiting for a spec sheet. A teaser is not a product, and a record-VRAM claim is not a benchmark. But the fact that a Chinese GPU company chose to lead with memory — and that the local-AI community immediately understood why — tells you exactly where the real bottleneck in consumer AI hardware sits right now.
Sources: the discussion on LINUX DO; Moore Threads' official MTT S80 specifications; Wccftech's Lushan coverage; Hardware Busters' teaser analysis; the U.S. Federal Register Entity List notice of October 19, 2023. For background on the domestic-GPU software stack, see our earlier piece on Tilelang, the Ascend alternative to CUDA, and on running models locally, Qwen3-8B running on a single GPU.