DeepSeek V4 & Huawei Ascend 950PR: Is Domestic Compute Ready for Enterprise AI?

Conclusion: Three facts and a warning

Fact 1: Ascend 950PR inference performance is 2.87× Nvidia H20, enabling a 70B model on a single card.

Fact 2: DeepSeek V4’s architecture efficiency breakthrough—CSA+HCA compresses KV cache to under 10% of the previous generation, making 1M‑token context cost acceptable.

Fact 3: V4‑Pro’s agent‑task score leads all open‑source models.

Warning: Hallucination rate of 94‑96% (higher than V3.2) stems from the compression trade‑off, making V4 unsuitable for high‑risk production tasks.

1. Ascend 950PR hardware overview

Key specs: FP4 compute 1.56 PFLOPS (unique), 112 GB HBM (17% more than Nvidia H20), memory‑access granularity 128 B vs 512 B (4× efficiency), single‑card 70B deployment (35 GB memory) versus multi‑card requirement for H20.

FP4 provides 4‑bit precision, giving four‑fold model density compared with FP16.

Memory‑access granularity reduces data movement four‑fold, benefiting long‑context KV cache.

Single‑card 70B deployment eliminates inter‑connect bandwidth cost.

2. DeepSeek V4 technical analysis

Model variants

V4‑Pro: 1.6 T parameters, 49 B activation, 1 M token context, 33 T training tokens, price 4 CNY per million tokens.

V4‑Flash: 284 B parameters, 13 B activation, same context and training data, same price.

CSA+HCA attention redesign

Traditional attention cost grows quadratically with context length; 1 M tokens require 64× the FLOPs of 128 K tokens.

CSA compresses every 4 tokens into 1 KV pair, selects important KV blocks via Lightning Indexer, and keeps a 128‑token sliding window for fine‑grained detail.

HCA further compresses every 128 tokens into 1 KV pair, enabling full dense attention on a much smaller base.

Measured on 1 M context, V4‑Pro uses 27 % of V3.2’s inference FLOPs, KV cache 10 % of V3.2, and KV cache is only 2 % of a BF16‑GQA8 baseline.

Other architecture upgrades

mHC (Manifold‑Constrained Super‑Connection): expands residual width 4×, uses Sinkhorn‑Knopp to constrain the gradient matrix to a doubly‑stochastic manifold, improving training stability at a cost of +6.7 % training time.

Muon optimizer: replaces AdamW with Newton‑Schulz iteration for gradient orthogonalization.

OPD‑after training: >10 domain experts perform KL‑distillation to stabilize agent training.

3. Empirical performance

Benchmark results

SimpleQA Verified: V4‑Pro 57.9 % (Claude Opus 46.2 %, GPT‑5.4 45.3 %).

Codeforces Rating: V4‑Pro 3206 (Claude Opus 3168, GPT‑5.4 3052).

SWE‑bench Verified: V4‑Pro 80.6 % (Claude Opus 80.8 %, GPT‑5.4 80.6 %).

Pythagorean Math 2025: V4‑Pro perfect 120/120.

V4‑Pro leads in three metrics and ties in two, the best open‑source performance compared with top closed‑source models.

Agent task ranking

V4‑Pro 1554 (open‑source #1).

Kimi K2.6 1484.

GLM‑5.1 1535.

MiniMax M2.7 1514.

V4‑Pro shows roughly a ten‑fold improvement over the previous generation and surpasses Gemini 3.1 Pro.

Ascend 950PR acceleration

General benchmark: 1.50–1.73× speedup.

Early‑version optimized benchmark: 35× over baseline (note: early version; production version lower).

4. Production‑grade drawbacks

Hallucination rate

V3.2: 82 %.

V4‑Pro: 94 %.

V4‑Flash: 96 %.

Higher compression leads to more fabricated answers, making V4 unsuitable for tasks demanding factual accuracy such as finance, healthcare, or law.

Token consumption cost

V4‑Pro consumes 190 million tokens in standard evaluation, costing $1,071.

Claude Opus 4.7 costs $4,811.

Kimi K2.6 costs $948.

DeepSeek V3.2 costs $71.

Although cheaper than top closed‑source models, the high token usage erodes cost advantage under high concurrency.

5. Developer guidance

Use‑case suitability

Ultra‑long document analysis (>200 K tokens): ★★★★★ – 1 M context + low KV cost.

Agent automation: ★★★★ – best open‑source, but high hallucination.

Serious production (finance/medical): ★★ – not recommended.

Simple chat / copywriting: ★★★ – cheaper but less stable than V3.2.

Research / academic inference: ★★★★★ – full open‑source, strong benchmarks.

Deployment options

API (OpenAI‑compatible):

deepseek-v4-pro    # OpenAI‑compatible endpoint
deepseek-v4-flash  # lower‑cost variant

Local Ascend deployment (vLLM‑Ascend): modify one line:

model = LLM("deepseek-v4", device="cuda")  # original
from vllm_ascend import LLM
model = LLM("deepseek-v4", device="npu")   # change only this line

ModelScope one‑click deployment:

from modelscope import pipeline
pipe = pipeline(task="text-generation", model="deepseek-ai/DeepSeek-V4", device_map="npu")

Important notes

Ascend 950PR batch production expected H2 2026; current tests use pre‑release nodes.

CANN ecosystem still maturing; some third‑party libraries lack Ascend support.

Training not supported on 950PR; training requires Ascend 950DT (expected Q4 2026).

6. Final assessment

Inference performance and cost efficiency are now competitive, making domestic compute “good enough” for many enterprise AI workloads, especially long‑context and agent tasks. However, training capability, ecosystem breadth, and high hallucination rates keep it from fully replacing Nvidia‑based solutions.