# DeepSeek V5 is planned to be pre-trained on Huawei Ascend 950 accelerators

> DeepSeek's next model generation, V5, is planned to be pre-trained on Huawei's Ascend 950 chips. Until now, Chinese hardware mostly ran DeepSeek models rather than training them.
>
> OpenCheese AI · 2026.11.04 · Web page: https://opencheese.dev/blog/deepseek-v5-is-planned-to-be-pre-trained-on-huawei-ascend-950-accelerators

DeepSeek's next generation of large language models, known as V5, is planned to be pre-trained on Huawei Ascend 950 accelerators. Until now, Chinese chips have mostly done the quieter work of running DeepSeek models that someone else's hardware had already finished training. Building the model on domestic silicon from day one is a different job, and a far more visible one.

## What is actually being claimed

The claim is short. DeepSeek plans to use the Ascend 950 line for pre-training V5. If that holds up, V5 would be the first major model from the company to go through pre-training on the Ascend platform from the start.

The word "if" matters. The report describes plans, not a finished training run. It gives no release date, no cluster size and no statement from either company, so what we have is an intention with a detailed spec sheet attached.

Still, the direction is clear. The report frames this as a potentially important step in widening the use of Chinese computing hardware for AI development, which is a polite way of saying that running a model and creating one are different tests of a chip.

## Inference is the easy part

Inference means executing a neural network that already exists. Pre-training means producing that network in the first place. A chip that does the first well has not necessarily earned a place in the second, and that gap is what makes the V5 plan worth watching.

It also explains why the Ascend 950 family comes as two chips rather than one.

## Two chips, two jobs

Huawei's Ascend 950 line has two models, the 950PR and the 950DT. Both are built on the third generation of the Da Vinci architecture and are meant for large AI models. They split the work differently.

| | Ascend 950DT | Ascend 950PR |
|---|---|---|
| Main purpose | Training, and the decoding stage of inference | The prefill stage of inference |
| Memory | 144 GB of Huawei's own HBM | 128 GB onboard memory |
| Memory bandwidth | 4 TB/s | 1.6 TB/s |
| Interconnect bandwidth | 2 TB/s | Not stated |
| Number formats named | FP8, MXFP8, MXFP4 | FP4 (in the compute figure) |
| Claimed compute | 4x the previous generation, MXFP4 against BF16 | 1784 TFLOPS in FP4 |

The 950DT is the one that matters for V5, since it is the model aimed at training. Its memory is Huawei's own HBM, a detail worth a second look, because it means the memory is also part of the domestic story and not only the processor.

The 950PR is the specialist. It targets the prefill stage of inference, and it trades some capacity and bandwidth for that role.

## Reading the 4x figure with care

The performance claim for the 950DT is that its MXFP4 throughput is four times that of the previous generation in BF16. That is a comparison between two different number formats, not the same format on two generations of silicon.

It may be perfectly true, and it may describe a real advantage in the lower-precision formats the chip supports. But it does not tell you how much faster the new chip is at the same task in the same format. The report offers no such like-for-like number, and the same goes for the 1784 TFLOPS FP4 figure for the 950PR, which is a stated capability, not a measured result.

## What the report leaves open

A few things are missing, and they are the things that would turn a plan into news:

- **Confirmation.** The report says V5 is planned for Ascend 950 pre-training, with nothing on whether it has begun.
- **Scale.** There is no word on how many accelerators would be involved or how they would be networked.
- **Results.** No benchmarks or training outcomes are given for any Ascend-trained DeepSeek model.
- **Timing.** No date for V5 appears anywhere.

Until some of that shows up, V5 on Ascend remains a plan, albeit one with unusually specific memory bandwidth figures.

## Related posts

- 2026.11.07 · [Inkling 2 Small ranks third for website generation on DesignArena](https://opencheese.dev/blog/inkling-2-small-ranks-third-for-website-generation-on-designarena.md): Thinking Machines Lab's 594 billion-parameter model scored 1340 in DesignArena's website-generation ranking. It is available through Tinker, with an Apache 2.0 weight release planned after external safety checks.
- 2026.10.31 · [DeepSeek and Moonshot reportedly trained new models on GPT-6 Astra outputs](https://opencheese.dev/blog/deepseek-moonshot-reportedly-trained-on-gpt-6-astra-outputs.md): A report citing unnamed sources says DeepSeek and Moonshot AI used distillation on GPT-6 Astra to build Kimi K3.1 Pro and DeepSeek V4.1-Flash. Neither company's benchmark claims have been independently confirmed.
- 2026.10.29 · [Canva Craft, the company's first image model, plans a design before it draws one](https://opencheese.dev/blog/canva-craft-first-image-model-plans-design-before-generating.md): Canva has released Canva Craft, an image model trained on human-made templates that plans layout and text before generating. It ranks fourth on Design Arena, behind three OpenAI models.