NeuroAI NEUROAINEUROAI.SITE
China AI, reported clearly
1,300+ verified stories on models, robotics, BCI and the business of AI.
Browse stories
ESC
Home · en

A domestic accelerator just took on a 2.8-trillion-parameter model

Daily AI/BCI industry briefing (EN edition, 2026-11-18).

2026-11-18 · 3720 words · NeuroAI
A domestic accelerator just took on a 2.8-trillion-parameter model

On November 17, Moore Threads said it has completed training adaptation for Kimi K3.1 Pro, the open-weight model released by Moonshot AI, on its MTT S5000 accelerator card.

The claim sits between two numbers: 2.8 trillion parameters, and the memory ceiling of a domestic card.

Training, not inference

According to Moore Threads, the work covered the native FSDP2 multimodal training path — model loading, shard scheduling and gradient synchronization.

The distinction matters. Inference loads weights and runs a forward pass. Training also has to hold gradients and optimizer states, and keep them synchronized across thousands of cards. One dropped card can send the job back to the last checkpoint.

Until now, domestic accelerators were mostly adapted for inference on models this size. When Moonshot released the weights on November 5, Huawei's Ascend announced same-day inference support through vLLM Ascend and SGLang. Public information on the training side has been much thinner.

A 22.4TB ledger before the first token

A 2.8-trillion-parameter model needs roughly 2.8TB of memory to hold its weights at FP8 precision.

Training adds Adam optimizer states. At the common estimate of 8 bytes per parameter, that is another 22.4TB. On 80GB cards, the optimizer states alone would occupy 280 cards, before gradients, activations and communication buffers.

That arithmetic is mine, based on public formulas — not Moore Threads' figure. The company has not disclosed how many cards it used, how long training ran, or what utilization it reached.

So "training adaptation completed" should be read narrowly: the path runs. How fast, and at what cost, is still open.

Bigger models hit the memory wall first

Kimi K3.1 Pro is a post-trained version of K3. It uses a mixture-of-experts architecture: 2.8 trillion parameters in total, about 104 billion activated per inference, a 1-million-token context window, and native image input.

Sparse activation saves compute, not storage. The full weight set still has to be split across thousands of cards, and once interconnect becomes the bottleneck, adding cards stops buying speed.

That is where high-bandwidth memory and interconnect bandwidth come in. Institutions estimate the HBM supply gap at 20% to 30%. A domestic card maker that wants to scale has to secure that memory first.

Why domestic cards, for this model

Open-weight models sell to a different buyer than closed APIs. A closed model sells access; an open model sells machines.

Anyone willing to download 2.8 trillion parameters most likely wants the model inside their own data center. Those buyers are the most sensitive to compute cost, and the most willing to try domestic hardware.

That is the point of Moore Threads' announcement: customers no longer have to train on imported cards and switch to domestic ones for inference.

The exam has changed

For two years the domestic compute story was about being usable — running inference, absorbing call volume. The exam now is whether the stack can carry a full training run without rolling back because one card failed.

One adaptation is far from that bar. But it moves the question from whether to how long.

For developers, the test is simple: wait for Moore Threads to publish training duration and utilization. If those two numbers are released, the link is real.

The part still unproven

Adaptation is an engineering milestone, not a benchmark result. What it establishes is that the software stack can load, shard and synchronize a model of this size; what it does not establish is throughput.

Two numbers would settle that: time to train a fixed token budget, and model FLOPs utilization. Neither is public. Until they are, the honest reading is that a domestic cluster can start such a job, not that it can finish one economically.

The buyers who care are the ones building domestic compute clouds. For them, a domestic card that can train is a product they can sell to model teams that are not allowed to buy abroad.

The gap between inference and training is where domestic stacks have been judged before. Inference adaptation takes weeks; training adaptation takes a cluster that stays up for the length of a run. Moore Threads has now claimed the second, at one model and one size. The claim will be tested by the next model, not this one.

*Disclaimer: This article is based on publicly available information and is for reference only. It does not constitute investment advice.*

Related coverage

Cite this article

NeuroAI (脑机网), "A domestic accelerator just took on a 2.8-trillion-parameter model," 18 November 2026, https://neuroai.site/a/01-na-model-domestic-card-2-8t-1118

Machine-readable: BibTeX · RIS · JSON · Markdown

More in “en” → · Back to home · Markdown version

China AI, reported clearly
1,300+ verified stories on models, robotics, BCI and the business of AI.
Browse stories