HK's DeepSeek Model Runs on China Chips Abroad
Sovereign LLM fully on China chips launches abroad—key for infra sovereignty tests.
30-Second TL;DR
What Changed
HKGAI-V3 based on DeepSeek V4 with full-parameter fine-tuning
Why It Matters
Advances China's AI independence by enabling advanced models on domestic chips, potentially lowering costs and geopolitical risks for users. Positions Hong Kong as a hub for exporting China-centric AI solutions globally.
What To Do Next
Monitor HKGAI site for HKGAI-V3 H1 release and benchmark on Chinese chips vs DeepSeek.
Key Points
- •HKGAI-V3 based on DeepSeek V4 with full-parameter fine-tuning
- •Optimized to run entirely on Chinese-made chips
- •Government-backed lab targeting sovereign AI exports
- •Planned unveiling in first half of this year
Deep Insight
Background and context from public sources — not the original article. 3 sources cited.
Enhanced Key Takeaways
- •DeepSeek V4, the underlying architecture for HKGAI-V3, was released on April 24, 2026, featuring a 1.6 trillion-parameter Mixture-of-Experts (MoE) design and a 1-million-token context window.
- •The initiative is part of a broader 'de-CUDA-fy' strategy in China, where DeepSeek provided exclusive early optimization access to Huawei and other domestic chipmakers, intentionally excluding Nvidia and AMD from pre-release hardware tuning.
- •HKGAI's sovereign AI approach focuses on 'governance-embedded' alignment, specifically addressing Hong Kong's unique multilingual (Cantonese, Mandarin, English) and socio-legal requirements under the 'one country, two systems' framework.
Competitor Analysis
- DeepSeek V4-Pro
- 1.6T MoE
- GPT-5.5
- Proprietary
- Gemini 3.1-Pro
- Proprietary
- DeepSeek V4-Pro
- 1M tokens
- GPT-5.5
- 1M+ tokens
- Gemini 3.1-Pro
- 1M+ tokens
- DeepSeek V4-Pro
- $1.74 / $3.48
- GPT-5.5
- $5.00 / $30.00
- Gemini 3.1-Pro
- N/A (Enterprise)
- DeepSeek V4-Pro
- Huawei Ascend 950
- GPT-5.5
- Nvidia H100/B200
- Gemini 3.1-Pro
- Google TPU
- DeepSeek V4-Pro
- 52 (Index)
- GPT-5.5
- 60 (Index)
- Gemini 3.1-Pro
- N/A
| Feature | DeepSeek V4-Pro | GPT-5.5 | Gemini 3.1-Pro |
|---|---|---|---|
| Architecture | 1.6T MoE | Proprietary | Proprietary |
| Context Window | 1M tokens | 1M+ tokens | 1M+ tokens |
| Pricing (Input/Output per 1M) | $1.74 / $3.48 | $5.00 / $30.00 | N/A (Enterprise) |
| Hardware Optimization | Huawei Ascend 950 | Nvidia H100/B200 | Google TPU |
| Benchmark (Reasoning) | 52 (Index) | 60 (Index) | N/A |
Technical Deep Dive
- Model Architecture: DeepSeek V4-Pro utilizes a 1.6 trillion-parameter Mixture-of-Experts (MoE) architecture with 49 billion active parameters per token.
- Training/Inference: Employs FP4+FP8 mixed-precision training. Inference is specifically optimized for Huawei Ascend 950 supernode clusters.
- Performance: V4-Pro achieves a single-card decode throughput of 4,700 tokens per second (TPS) on Ascend 950 hardware for 8K input scenarios.
- Alignment: HKGAI-V1 (predecessor) utilized full-parameter fine-tuning to instantiate regional values, using bespoke benchmarks like HKMMLU (local knowledge), SafeLawBench, and Adversarial HK Value Bench.
Future ImplicationsAI analysis grounded in cited sources
Timeline
- 2025-01DeepSeek releases R1 model, sparking international attention for cost-effective performance.
- 2025-12DeepSeek releases V3.2 and V3.2 Speciale models.
- 2026-03HKGAI rolls out tools for school selection and budgeting as part of its governed AI agent network initiative.
- 2026-04DeepSeek releases V4-Pro and V4-Flash, with exclusive early optimization for Huawei Ascend 950 chips.
Sources (3)
Factual claims are grounded in the sources below. Forward-looking analysis is AI-generated interpretation.
Weekly AI Recap
Read this week's curated digest of top AI events →
Related Updates
AI-curated news aggregator. All content rights belong to original publishers.
Original source: SCMP Technology ↗
This is a summary, not the original. Read the source, or get the weekly briefing.
The weekly digest
One email a week. Unsubscribe anytime.