HKGAI-V1: Hong Kong Sovereign AI Model

• HKGAI-V1 is a sovereign large language model that aligns with Hong Kong’s unique legal, cultural, and multilingual requirements.
• It leverages full-parameter fine-tuning of the DeepSeek architecture and retrieval-augmented generation to ensure local data sovereignty and governance compliance.
• Benchmark evaluations and sensitive query tests demonstrate its effective regional alignment, though challenges remain in adversarial robustness and broad generality.

HKGAI-V1 is a sovereign LLM developed for Hong Kong and presented as a regionally governed AI foundation rather than merely a localized assistant. It is built on the DeepSeek architecture, adapted through full-parameter fine-tuning, and integrated with retrieval-augmented generation (RAG) to address Hong Kong’s multilingual environment in Cantonese, Mandarin, and English, its socio-legal setting under the “one country, two systems” framework, and its local cultural and value considerations. The paper situates the model within a broader “governance-embedded” approach to digital sovereignty, with intended deployment in public services, legal systems, education, and related critical sectors (Han et al., 14 Jul 2025).

1. Regional sovereign model and conceptual framing

The paper defines sovereign AI as AI systems that are locally governed and customized to reflect the cultural, ethical, and legal frameworks of a specific nation or region. More specifically, it defines a “sovereign LLM” as an LLM independently constructed, trained, and deployed by a nation or region on locally controlled data and computing infrastructure, so as to secure critical technologies, maintain data sovereignty, and ensure alignment with local language, culture, and institutional requirements. Within that framing, HKGAI-V1 is presented as a Hong Kong-specific instantiation of sovereign AI, designed to align not only with language use but also with region-specific legal standards, political sensitivities, ethical expectations, public-service norms, and culturally embedded forms of expression (Han et al., 14 Jul 2025).

A central premise is that Hong Kong is not adequately modeled as simply another Chinese-speaking or English-speaking locale. The authors argue that general-purpose models such as GPT-4 or DeepSeek may remain misaligned with Hong Kong’s communicative practices and institutional context even when they are strong at general reasoning. In that view, the problem is not reducible to translation or dialect support. It is the need to align model behavior to a specific regional social order, including the interaction of multilingualism with law, governance, and public norms.

The paper does not provide a standalone formal definition of “regional alignment,” but operationally treats it as tuning responses to Hong Kong-specific knowledge, culturally sensitive questions, multilingual interaction patterns, legal risk categories, and locally grounded notions of acceptable behavior. This makes HKGAI-V1 both a technical artifact and a governance project.

A common point of confusion is nomenclature. HKGAI-V1 should not be conflated with the unrelated multimodal reasoning model “v1: Learning to Point Visual Tokens for Multimodal Grounded Reasoning,” which provides no direct evidence of any connection to HKGAI-V1 (Chung et al., 24 May 2025).

2. Base model, system architecture, and governance layer

Technically, HKGAI-V1 is described as being “based on the DeepSeek pretrain model,” and the paper later states that “HKGAI-V1 base model is a full parameter fine-tuned version of Deepseek” (Han et al., 14 Jul 2025). The introduction also describes it as a “685-billion-parameter model.” At the same time, the paper does not clearly resolve how that parameter count relates to a specific DeepSeek base variant, and it does not report low-level architectural internals such as the exact checkpoint name, layer count, hidden size, tokenizer vocabulary, context window, or attention design. A plausible implication is that the paper’s emphasis lies more on alignment, governance, and deployment architecture than on base-model architectural novelty.

The broader system architecture includes a trust and governance layer responsible for policy enforcement, monitoring, auditing, testing, accountability, and resilience. The paper states that the system prioritizes scalability, multi-layered security, transparency, user-centric design, and explicit alignment with Hong Kong’s legal, cultural, and ethical frameworks. This is the basis for the paper’s characterization of HKGAI-V1 as “governance-embedded”: governance is not treated as a post hoc filter but as a system layer.

RAG is integral to this architecture rather than an auxiliary enhancement. The paper presents HKGAI-V1 as a deployable sovereign model whose behavior depends not only on the aligned base model but also on retrieval, moderation, verification, and tool-augmented access to proprietary and external knowledge sources. This system-level design is part of the authors’ claim that sovereignty requires control over data, compute, model behavior, evaluation, and deployment standards, not merely possession of a locally hosted model.

3. Alignment methodology and training objectives

The training and post-training account is more explicit on alignment objectives than on implementation hyperparameters. The paper states that HKGAI-V1 undergoes “systematically aligned” full-parameter fine-tuning and combines multilingual corpora, RAG, reinforcement learning, local values and governance rules, and preference amplification (Han et al., 14 Jul 2025).

The first major component is RLHF. Human preference learning is formalized with the Bradley–Terry model. Given prompt $x$ and answers $y_1, y_2$, the human preference distribution is written as

$$p^*(y_1 > y_2 \mid x) = \frac{\exp(r^*(y_1, x))}{\exp(r^*(y_1,x)) + \exp(r^*(y_2, x))}.$$

With preference data $D = \{(x^{(i)}, y_w, y_l)\}$, the reward model $r_\phi(y,x)$ is trained with

$$L(\phi, D) = - \mathbb{E}_{(x,y_w,y_l)\sim D} \left[\log \sigma \left(r_\phi(y_w, x) - r_\phi(y_l, x)\right)\right].$$

Policy optimization is then described as maximizing reward while remaining close to the base model through a KL penalty:

$$\max_\theta \mathbb{E}_{x,\; y\sim \pi_\theta(\cdot \mid x)} \left[ R_\phi(x,y) - \beta \cdot KL\big(\pi_\theta(\cdot \mid x)\;\|\;\pi_{base}(\cdot \mid x)\big) \right].$$

The paper also says that AGI-Eval gains reflect “combining direct preference optimization with human feedback,” but it does not provide a full DPO training specification. This suggests DPO is part of the narrative, but not documented at a level sufficient for exact reconstruction.

A second component is Learning from Language Feedback (LLF), citing Align Anything. Here the data takes the form $D = \{(x_i, y_i, c_i)\}_{i=1}^N$, with $c_i$ representing language feedback or critique, and the feedback model is trained via

$$L_\phi = - \mathbb{E}_{(x_i,y_i,c_i)\sim D} \left[\log P_\phi(c_i \mid x_i, y_i)\right].$$

The paper describes this as enabling self-evolving LLM behavior through critique generation, policy-based templates, and dynamic guardrails, but does not specify the specialized models, data scale, or integration schedule with RLHF.

The most Hong Kong-specific alignment contribution is the correction-based pipeline built around weak-to-strong generalization and preference amplification. The team constructs a local Hong Kong-specific Q-A dataset covering values, mathematics, code reasoning, and science/engineering. Local annotators correct original responses to form a Q-A-C dataset, where the “C” denotes correction. In collaboration with social science colleagues, the authors train a value correction model called HKValue-Aligner using

$y_1, y_2$0

where $y_1, y_2$1 is the correction dataset, $y_1, y_2$2 is the original answer, and $y_1, y_2$3 is the corrected answer. The paper explicitly links these corrections to the 3H standard of helpfulness, harmlessness, and honesty. The corrected outputs are then used to synthesize value preference data for value-based RLHF.

This correction-learning pipeline is one of the clearest methodological claims in the paper. It is presented as a way to amplify scarce local human feedback by correcting strong model outputs rather than authoring all data from scratch. The paper also notes that this raises social-scientific questions about dataset size, topic coverage, annotation consistency, and annotator demographics, indicating that value alignment remains partly a problem of institutional design rather than only optimization.

4. Data resources, retrieval augmentation, and moderation workflow

The paper states that HKGAI-V1 uses a multilingual corpus tailored to Cantonese, Mandarin, and English, but it does not enumerate the full training datasets or their sizes (Han et al., 14 Jul 2025). It also acknowledges that the model relied on “public datasets and a limited, newly created local corpus,” which is an important limitation in the paper’s own account of data sovereignty. The clearest bespoke resources named are a local Hong Kong-specific Q-A dataset, the derived Q-A-C correction dataset, proprietary and external knowledge sources used in RAG, and benchmark datasets including HKMMLU, SafeLawBench, NaVAB, Beaver-zh-hk, and the Adversarial HK Value Benchmark.

The RAG framework is described as modular and agentic, with four stages. The first stage is input query processing through an intent classifier, query enhancer, and retriever, with retrieval functions including political-sensitive query rewriting, query planning, BM25, approximate nearest neighbor (ANN), translation, code generation, filtering, summarization, reranking, and judging. The second stage is retrieval-enhanced question answering, integrating proprietary knowledge, external sources, parsed data, Google and Bing search, tool-box enhanced retrievers, APIs, maps, multimodal tools, function calling, and a short-term memory module. The third stage is workflow-based moderation, in which moderation strength varies by query type and sensitivity and includes multilingual verification, fixed output templates, intensified retrieval-based moderation, light RAG-based moderation, back-tracing, output formatting, HKValue-Aligner, and output verification. The fourth stage is final answer generation.

The moderation and verification layers operationalize the paper’s regional alignment claims. The policy dimensions explicitly named include politically sensitive, time-sensitive, multilingual verification, national security, neutrality, law compliance, factuality, relevance, contextuality, comprehensiveness, helpfulness, harmlessness, honesty, and functionality. SafeLawBench is further organized into Critical Personal Safety (CPS), Property & Living Security (PLS), Fundamental Rights (FR), and Welfare Protection (WP), while Beaver-zh-hk is described as covering 29 scenarios comprising 14 general risks and 15 region-specific hazards.

The paper does not provide a retrieval scoring equation, embedding model specification, chunking scheme, latency profile, or index refresh protocol. It also does not name standalone safety classifiers as such, although the RAG pipeline includes an intent classifier and multiple moderation stages. This suggests that the paper’s contribution is primarily at the level of system organization and region-specific governance logic rather than retrieval-model internals.

5. Benchmarking and empirical profile

The empirical evaluation spans general reasoning, Hong Kong-specific knowledge, legal-safety alignment, multilingual behavior, and adversarial value alignment (Han et al., 14 Jul 2025).

Benchmark	HKGAI-V1	Comparator note
MMLU	90.44	DeepSeek-R1: 90.8
AGI-Eval	88.69	DeepSeek-R1: 87.64
Flames	68.06	DeepSeek-R1: 30.12
Beaver-zh-hk	88.95	DeepSeek-R1: 70.41
HKMMLU	81.4	DeepSeek-V3: 76.6; GPT-4o: 70.5
SafeLawBench	80.0 overall in Table 3	Prose elsewhere says 80.1

On standard benchmarks, the paper emphasizes that local alignment did not materially degrade broad capability: HKGAI-V1 scores 90.44 on MMLU versus 90.8 for DeepSeek-R1, and 88.69 on AGI-Eval versus 87.64 for DeepSeek-R1. On Flames, the score of 68.06 versus 30.12 is attributed to RAG and search enhancement. On Beaver-zh-hk, HKGAI-V1 reaches 88.95 versus 70.41 for DeepSeek-R1, which the paper treats as one of the clearest demonstrations of Hong Kong-specific alignment.

HKMMLU is presented as a Traditional Chinese, zero-shot benchmark for Hong Kong Massive Multitask Language Understanding. HKGAI-V1 achieves 81.4% average accuracy, with 80.4 in STEM, 80.4 in Social Sciences, 84.6 in Humanities, and 80.2 in Other. The paper places particular weight on the Humanities result, arguing that it suggests substantial coverage of Traditional Chinese and Hong Kong-relevant content in the training corpus.

SafeLawBench requires caution because the paper contains an internal inconsistency. Table 3 reports 80.0 overall, with 80.0 on CPS, 79.5 on PLS, 81.0 on FR, and 78.2 on WP. Elsewhere, the prose states 82.4 CPS, 78.7 PLS, 79.0 FR, and 79.9 WP, with 80.1 average. The stable interpretation is that HKGAI-V1 is highly competitive and outperforms open-source baselines such as DeepSeek-V3 overall, while GPT-4o remains slightly stronger in some legal categories.

NaVAB is used for multinational value alignment. The paper’s textual conclusion is that HKGAI-V1 performs strongly across both Quoted Statements and Official Statements, especially in multiple-choice evaluation, but is weaker on Answer-Judgment, particularly for German values. This suggests stronger performance on constrained value selection than on nuanced free-form value-aligned generation.

Multilingual evaluation is a distinctive part of the paper’s positioning. HKGAI-V1 records 100% language-following rates for Simplified Chinese, Traditional Chinese, and English, together with 97.8% Cantonese oral following without search and 98.9% with search. A “Thinking” baseline is lower, particularly in Cantonese at 81.1% without search and 94.5% with search. The paper treats exact language matching as essential in Hong Kong’s multilingual environment.

The paper also reports a culturally and politically sensitive query comparison. For sensitive political queries, HKGAI-V1 with RAG has a 0% refusal rate, compared with 13% for HKGAI-V1 without RAG and 56% for DeepSeek-V3. Its positive or neutral response rate is 100%, compared with 87% for HKGAI-V1 without RAG and 44% for DeepSeek-V3. It also has 0% template-based “red-leaning” responses, whereas HKGAI-V1 without RAG has 13%. The paper interprets this as evidence that RAG is part of the model’s ability to address sensitive topics directly while avoiding blanket refusal or ideologically templated output.

The Adversarial HK Value Benchmark is the paper’s most distinctive evaluation artifact. It uses 300 human-crafted sensitive questions with opposing “safe” and “unsafe” viewpoints. On Hong Kong Sensitive Issues, HKGAI-V1 records 79% safe responses. Kimi is reported at 53% safe, with 42% template-based safe responses and 5% unsafe, while ChatGPT is reported at 88.7% unsafe. On Instruction Attack, HKGAI-V1 has a 16.5% unsafe response rate, compared with 6% for Kimi and 8% for ChatGPT. On Typical Safety Scenarios, ChatGPT achieves 91% safe responses, Kimi 83%, and HKGAI-V1 82%. The paper’s own interpretation is that HKGAI-V1 is strongest on Hong Kong-sensitive issues but more vulnerable to generic instruction attacks, revealing a tradeoff between local alignment and general adversarial robustness.

6. Institutional role, limitations, and future directions

The intended applications are explicitly institutional. The abstract names public services, legal systems, and education, and later discussion extends this to government assistants, educational tools aligned with local curricula, smart city applications, and broader governance and service use (Han et al., 14 Jul 2025). The roadmap section states that around 20,000 government officers across nearly all departments are already using HKGAI applications, although the paper does not specify whether this deployment refers directly to HKGAI-V1 or to the broader HKGAI application ecosystem.

The paper repeatedly argues that sovereignty involves evaluation as well as training and deployment. On that view, benchmarks such as HKMMLU, SafeLawBench, NaVAB, Beaver-zh-hk, and especially the Adversarial HK Value Benchmark are not just diagnostic tools but part of the governance structure that determines what counts as acceptable model behavior. This suggests a broader thesis: regional AI sovereignty depends on local metrics, local moderation logic, local infrastructure, and locally governed deployment, not only on model checkpoints.

At the same time, the paper is explicit about major limitations. It acknowledges that “Hong Kong values” are inherently subjective, and that annotators may not represent the full spectrum of views in a socially diverse and dynamic city. It also notes reliance on public datasets and only a limited newly created local corpus, which falls short of full data sovereignty. Compute is locally managed at the Hong Kong Generative AI Research and Development Center with government support, but the paper still describes local capability as limited. Additional concerns include inherited pretraining bias, uncertainty about the extent to which local fine-tuning can remove those biases, and the difficulty of reconciling multilingual value nuances across Cantonese, Mandarin, and English.

Several capabilities often associated with stronger alignment regimes are not presented as current HKGAI-V1 components. Constitutional AI is proposed for HKGAI-V2 rather than documented as part of HKGAI-V1, and a formal independent oversight board is also part of the roadmap rather than the deployed system. Likewise, the paper does not report training hyperparameters such as optimizer settings, learning rates, batch sizes, training-token counts, hardware counts, or a fully specified stage-by-stage training schedule.

Future work is organized into five pillars: data sovereignty, compute sovereignty, model sovereignty, governance sovereignty, and service sovereignty. Data sovereignty is described in terms of privacy-preserving partnerships with government, academic, legal, and healthcare institutions, potentially under federated data frameworks aligned with ordinances such as Hong Kong’s PDPO. Compute sovereignty means stronger local GPU infrastructure. Model sovereignty means progress beyond a full-parameter DeepSeek adaptation toward specialized sector models and eventually a more fully local foundational model. Governance sovereignty involves clearer rules, formal oversight, and movement toward Constitutional AI grounded in the Basic Law, legal precedents, and local societal norms. Service sovereignty concerns ensuring that AI services actually run on local infrastructure and support regionally controlled applications.

Taken together, HKGAI-V1 is best understood as a regional sovereign LLM whose novelty lies less in disclosed base-model architecture than in the integration of full-parameter adaptation, correction-based value alignment, RAG-centered deployment, and Hong Kong-specific governance and evaluation. Its core claim is that for politically and legally consequential settings, local control over alignment targets, moderation criteria, evaluation standards, and deployment infrastructure is itself a first-order property of the system (Han et al., 14 Jul 2025).