DeepSeek plans to train V5 model on Huawei Ascend chips

Beijing-based AI company HongYun has unveiled HyperLM V2.5, a new version of its platform for combined training and inference of large models. This was reported by IT Home and ifeng Tech on November 3rd. The company claims to be the first to achieve deep adaptation to the Huawei Ascend 950DT chip, full compatibility with the air-cooled Atlas 850E SuperPoD system, and successful technical certification from Huawei for Ascend Native.

HongYun believes that defining support is often too vague, where a model is only considered supported after installation and token generation. The company aims to assist government and enterprise clients using private deployments in achieving maximum performance from new accelerators.

The company describes the Ascend 950DT as a flagship chip for high-throughput tasks such as extended chat interactions, processing long documents, and continuous agent reasoning. Furthermore, it is noted that the air-cooled Atlas 850E system can be integrated into existing data centers without requiring infrastructure upgrades.

HyperLM V2.5 adapts at multiple levels. At the interconnect level, the platform supports Huawei's UnifiedBus 2.0 technology, which provides a system interconnect bandwidth of 13.4 TB/s, preventing bottlenecks during traffic transfer between cards—a crucial factor for long contexts or splitting prefill and decoding stages. At the operator level, the system is optimized for mxFP4 inference, enabling 14.2 PFLOPS per node.

The platform supports the latest open-source models, including Qwen, GLM, and DeepSeek, and its scheduler has been rebuilt considering the memory profile and throughput of the 950DT.

The release of HyperLM V2.5 includes five new features. Prefill and decoding splitting allows the two stages to be isolated into independent resource pools so they do not compete during peak loads. Multi-level KV caching reuses system prompts, frequently encountered contexts, and fixed session templates to reduce redundant computations. vNPU virtualization segments the Ascend 950DT, 910C, and 910B chips, providing granular quotas and isolation for AI Core computation and HBM memory, and also supports hot migration and automatic rescheduling upon hardware failure.

A unified gateway model, compatible with the OpenAI protocol, has also been added, which manages routing, rate limiting, beta releases, and A/B testing. Comprehensive monitoring across the entire stack tracks load, memory usage, throughput, latency, and QPS, providing alerts and automated fault localization.

For regulated industries such as government administration, energy, transportation, and finance, the platform provides role separation for administrator, security officer, and auditor, multi-level classification control, complete operation logs, and tracking of computational resources, model call audits, and usage statistics per project.

HongYun was incubated by the Western Institute of Computing Technology of the Chinese Academy of Sciences, and its core team consists of employees from the Institute of Computing Technology and Sugon. In 2020, the company joined Huawei's Ascend partnership program. The company plans to continue development based on Ascend and Hygon DCU hardware, focusing particularly on heterogeneous scheduling, inference acceleration, caching, and agent applications.

Developers discovered three new model identifiers in DeepSeek's official Harness repository on GitHub: deepseek-v5, deepseek-v5-open, and deepseek-v5-next. Currently, DeepSeek has not announced the V5 model, and these entries are test placeholders in the code, not available models.

Upon checking the repository, it was determined that these identifiers are located in a test file for the DeepSeek account provider, specifically in the file packages/llm/llm-deepseek-account/tests/provider.spec.ts, on lines 154–156, next to the existing deepseek-v4-pro entry. The repository history shows they were added in a commit from September 24th, which moved the DeepSeek account model directory into the official list of models.

The test scenario checks how the client filters models by required version: the v5-open entry requires a minimum client version of 0.0.1, while v5-next requires 99.0.0-alpha.1, and the test expects only the flash and v5-open models to appear in the list. This indicates the possibility of switching to a new model from the Harness server side, either for all clients or only for those who have updated, without needing a new release. However, this does not confirm the existence of the V5 model or its release timeline, and some developers consider these names merely test placeholders. As of October 28th, DeepSeek has made no announcements, model cards, or API identifiers for the V5 series.

The current model lineup includes V4-Pro and V4.1-Flash; the latter was released in September with a new architecture while retaining the V4 name.

In the same week, a new benchmarking result was presented for the current model. Artificial Analysis released Intelligence Index v5, which is the largest update since the V4 release in January, with 60% of the weight now based on closed test datasets or hidden answers. Coding accounts for 25% of the index, and of these 15 points, FrontierCode 2.0 from Cognition accounts for the evaluation of code changes in 132 real open-source tasks to see if they would approve a pull request merge.

On the FrontierCode 2.0 platform, the DeepSeek V4.1-Flash-1017 model scored 28% at maximum effort, which is the highest score among models with released weights. Kimi K3.1 Pro achieved 25.5%, but its full weights are not yet available. Gemini 4 Argon reached 25.7%, and GPT-6 Luna scored 25.2%.

The overall index was less favorable. DeepSeek V4.1-Flash scored 32 points, tying with GLM-5.3, while the leader, Claude Opus 5.5, received 58 points. Artificial Analysis notes that the gap between the best open model with released weights and the leaderboard leader has increased from 3 points under v4.3 to 12 points under v5.

DeepSeek has released DeepInfer, an inference execution framework that forms the basis of its large model services. This release occurred on October 20th.

DeepInfer supports Nvidia GPUs as well as Huawei Ascend processors. The framework provides several features, including distributed inference, speculative decoding, and structured output.

According to DeepSeek's statement, the goal of this release is to help developers and enterprises increase throughput and reduce latency while decreasing computational costs and energy consumption.

Furthermore, the company made the deepseek-recipe component publicly available. It is responsible for transforming API requests and handling model input and output in accordance with DeepSeek's proprietary services. Currently, DeepInfer supports the DeepSeek-V4.1-Flash-1017 model, and the company intends to continue releasing inference code and optimizations for future models through this project.