Huawei Ascend servers
Atlas 800I A2 with eight Ascend processors and up to 512 GB of memory, and Atlas 650E with eight Ascend 950DT and 768 GB, for large models such as DeepSeek and GLM.
Chinese AI hardware: Huawei Ascend and other Chinese accelerators
Chinese AI accelerators now run large models: vLLM runs DeepSeek-V4 and GLM-5 on Huawei Ascend, and Cambricon and Moore Threads have announced support for new DeepSeek releases on the day they shipped. We supply and set up Ascend cards and servers, other Chinese accelerators and high-memory Chinese mini PCs.
The software picture, plainly
Chinese hardware has to be judged together with its software. Huawei Ascend runs on CANN with the official vllm-ascend plugin and serves DeepSeek-V4, GLM-5 and Qwen3. Cambricon and MetaX have their own vLLM plugins, and Moore Threads is the only other Chinese vendor llama.cpp officially supports. Ollama officially runs on none of them. That is why we test your model on the same card before you buy.
Which tier for which job
An inference card in an ordinary server
Atlas 300I Duo with 96 GB and Atlas 350 with 112 GB of memory, in ordinary PCIe servers, for mid-size models.
Other Chinese accelerators
Moore Threads, Cambricon, MetaX and Hygon, each with the software and models officially tested on it.
High-memory Chinese mini PCs
GMKtec, Beelink and Minisforum build small computers on AMD Ryzen AI Max with 128 to 192 GB of memory, for a department's assistant and MoE models.
Huawei Ascend cards and servers
Specifications from Huawei's official pages and guides, reviewed on 22 October 2026.
| Product | Processor | Memory | Bandwidth | What fits |
|---|---|---|---|---|
| Atlas 300I Pro | One Ascend 310P | 24 GB LPDDR4X | 204.8 GB/s | A model of about 8B at 8-bit; for one or two users |
| Atlas 300I Duo | Two Ascend 310-series | 48 or 96 GB LPDDR4X | 408 GB/s | With 96 GB, a model of about 30B at 8-bit; Qwen3-30B-A3B is officially supported in vllm-ascend |
| Atlas 350 | Ascend 950PR | 112 GB HBM | 1.4 TB/s | A model of about 70B at 8-bit on one card |
| Atlas 800I A2 (server) | Eight Ascend 910-series | Eight x 32 or eight x 64 GB, up to 512 GB | Not published | With the 64 GB version, an MoE of about 235B at 8-bit; 671B-class models need two servers |
| Atlas 650E (server) | Eight Ascend 950DT | Eight x 96 GB HBM, 768 GB in total | Up to 4 TB/s per processor | 671B-class MoE models at 8-bit (tight); a 14U rack unit drawing 14.5 kW |
| Atlas 950 SuperPoD | Up to 1,024 Ascend 950DT | 96 GB per processor | Up to 4 TB/s per processor | A liquid-cooled data-centre cluster for trillion-parameter models |
| Orange Pi AI Studio and AI Studio Pro | Ascend 310-series, 176 and 352 TOPS | 48 to 192 GB LPDDR4X | Not published | A large model fits but writes slowly; a desktop box with a single USB4 port |
The official vllm-ascend support table covers the A2 and A3 series, 950DT products and the Atlas 300I Duo (experimental), but not the Atlas 200I A2. The "what fits" column is our estimate: each billion parameters takes about 1 GB of memory at 8-bit.
Other Chinese accelerators
Specifications from the vendors' official pages, reviewed on 22 October 2026. Where a vendor has not published a figure, the table says so.
| Product | Memory | Bandwidth | Software | Status |
|---|---|---|---|---|
| Moore Threads MTT S4000 | 48 GB | 768 GB/s | MUSA; llama.cpp and SGLang officially, vLLM through the vendor's vllm-musa | A model of about 30B at 8-bit per card |
| Moore Threads MTT S5000 | Not published by the vendor | Not published by the vendor | vllm-musa, with serving recipes for DeepSeek-V4-Flash and Qwen | The vendor announced day-one support for Kimi K3, GLM-5.3-Flash and DeepSeek-V4.1-Flash |
| Cambricon MLU370-X8 and MLU370-X4 | 48 and 24 GB LPDDR5 | 614 and 307 GB/s | NeuWare; the vllm-mlu plugin and LMDeploy | The vendor announced day-one DeepSeek-V4 support; newer generations have no published specifications |
| MetaX C500, C550 and C588 | 64, 64 and 128 GB | Not published by the vendor | MXMACA; the vLLM-metax plugin and LMDeploy | C550 and C588 are OAM modules for eight-way servers |
| Hygon DCU (K100-AI) | Not published by the vendor | Not published by the vendor | DTK, derived from ROCm; vLLM through the vendor's container images | Only after testing your model on the same card |
| Kunlunxin P800 (Baidu) | Not published by the vendor | Not published by the vendor | The vLLM-Kunlun plugin | Only after testing your model on the same card |
None of these cards officially supports Ollama. Smaller vendors such as Biren, Enflame, Iluvatar CoreX and T-Head exist too, but model-serving software for them is more limited. We also supply servers from Chinese makers such as Inspur, H3C, Lenovo WenTian and xFusion, with these cards or with NVIDIA and AMD cards.
Chinese mini PCs with Ryzen AI Max
Small computers that Chinese makers build on AMD processors with a lot of shared memory. Specifications from each maker's official store, reviewed on 22 October 2026.
| Machine | Processor and memory | Network and ports | Suits |
|---|---|---|---|
| GMKtec EVO-X2 and EVO-X3 | Ryzen AI Max+ 395, up to 128 GB | USB4 and 2.5GbE; EVO-X3 adds OCuLink | A department's assistant and MoE models |
| Beelink GTR9 Pro | Ryzen AI Max+ 395, 128 GB | Two 10GbE and two USB4 ports | Joining several machines over fast networking |
| Minisforum MS-S1 MAX | Ryzen AI Max+ 395, 128 GB | Two 10GbE, USB4 v2 and a PCIe x16 slot; a rack kit for four units | Several machines in one rack |
| GMKtec EVO-X5 Pro and Minisforum MS-S1 MAX-P495 | Ryzen AI Max+ PRO 495, 192 GB (up to 160 GB for the GPU) | Two 10GbE and USB4 v2 | Larger models on one small machine |
Ryzen AI Max processor specifications are on the AMD page.
From the first call to a working system
Hardware is chosen after a model has been measured on your own work, not from a table. A quote is issued only once it is clear which model has to serve how many people, and how fast.
- 01
First conversation
We go through the task, the number of users, the data, where the hardware goes and what you already own.
- 02
Measured on your work
We test several open-weight models on samples of your real work and size memory and speed from those results.
- 03
Written quote
You get the exact configuration, the proposed model, the delivery time and the support terms in writing. Prices appear only in the quote.
- 04
Installation
We install the hardware, bring up the model and connect the assistant to your systems, then test it once with international internet cut off.
- 05
Handover and support
You get documentation and training for your team; afterwards we maintain the system and update the models under contract.
How to get in touch
Fill in the form or message us on WhatsApp. The clearer these are, the more exact the quote:
- What the model has to do (answer from documents, code, summarise, process images or speech)
- How many people use it at the same time, and what response time is acceptable
- The model or hardware family you have in mind, if you already know
- Where it goes: your server room, a data centre, or hardware you already own
- How sensitive the data is, and whether it must run with no internet at all
We reply within one business day.
Zharf
Try the ZharfAI assistant right now
Ask about Persian assistants, document automation, or how to start a pilot.
Frequently asked questions about Chinese AI hardware
Can Huawei Ascend replace NVIDIA?
For serving models, in many cases yes: vLLM runs on Ascend through the official vllm-ascend plugin, with official support for DeepSeek-V4, GLM-5 and Qwen3. But NVIDIA's tooling is broader, some models and features reach Ascend later, and Ollama does not officially run on it. That is why we test your model on the same hardware before you buy.
Which models run on Huawei Ascend?
According to the official vllm-ascend table: DeepSeek-V4-Flash and DeepSeek-V4-Pro (on Ascend 950DT and the A2 and A3 series), GLM-5.1 to GLM-5.3, and Qwen3, whose Qwen3-30B-A3B also runs on the Atlas 300I Duo. DeepSeek itself has also confirmed DeepSeek-V3 on Ascend.
Does Xiaomi make AI server hardware?
Not yet. Xiaomi makes phone and car chips, and announced its first AI accelerator (XRING O100) in August 2026 for running models on the device itself; reports say it becomes commercial in 2027. Xiaomi's open MiMo models, however, are available under the MIT licence, and we run them on your hardware.
Why are some specifications missing from the table?
Because the vendor has not published them. Hygon has no public datasheet, and MetaX and Moore Threads do not state memory or bandwidth for some products. We left out figures that appear only in unofficial reports, and we measure each card's real capability by testing your model on it.
Do you rent GPU servers by the hour or month?
No. ZharfAI does not rent GPUs by the hour or month. There are three ways to work with us: we supply the hardware and install it on your premises, we run a dedicated machine for you and maintain it ourselves, or we set the model up on hardware you already own. If you only need a few hours of compute for a test, say so in the first conversation and we will suggest a suitable route.
How much does it cost?
We do not publish prices, because they depend on the configuration, quantity, the day's market and how it is delivered. After the first conversation and a test on your work you receive a written quote that also names the exact hardware model, the delivery time and the support terms.
Does it work without international internet?
Yes, if the model, the serving software, the assistant, user sign-in and monitoring all sit inside your network. Anything that depends on a foreign service stops during a cut, so before handover we test the system with international internet deliberately cut off.
Test the model on Chinese hardware before you buy
Tell us which model you want to run and for how many users. We measure it on Ascend or another Chinese card, and if different hardware suits your work better, we will say so plainly.