Tencent released Hy4 preview on 28 August 2026. Its official materials describe a mixture-of-experts model with 770 billion parameters in total, 49 billion activated per token and a context window exceeding one million tokens. The more philosophically relevant fact is not the scale alone. Tencent also released model weights under the Apache 2.0 licence, independent deployment paths, and workflows for both full and LoRA fine-tuning. Reuters confirmed that this is an early preview intended for software engineering, research and financial analysis. An institution, team or individual with sufficient compute can therefore copy the same base weights, deploy them in different environments, attach different memories and tools, and continue training them on different data.
This raises a question more basic than whether the model is intelligent. Are systems copied from the same weights, and still described by the same model name, one thing? More precisely, can they be understood as one continuing subject? Ordinary language readily treats a model name as if it were a person's name. We say “Hy4 answered”, “Claude refused” or “the model remembers me”, as though one unchanging entity always stood behind the name. Open weights expose the ambiguity. One set of weights can generate many isolated processes, and those processes can then acquire different histories.
At least four levels need to be separated. A model lineage is the technical family formed by an architecture, base weights, training methods and version relationships. A running instance is a process executing on particular computational resources at a particular time. A persona is the recognisable presentation jointly produced by system prompts, interfaces, conversation history and product design. A subject requires stronger conditions: a boundary maintained across time, ownership of its own history, memory that causally shapes later states, and some continuing relation to itself. These levels can be related, but a shared name does not collapse them into one.
Two unmodified copies may have exactly the same parameters at the instant of copying, but qualitative identity is not numerical identity. They run on different hardware, receive different inputs, form different caches and conversation records, and may receive different tool permissions. An interaction experienced by instance A does not automatically become part of instance B's history. If B is later fine-tuned, it remains genealogically related to the base model while diverging further in behavioural disposition and conditions of use. Common origin can explain where the systems came from. It cannot by itself establish that they remain one subject.
Traditional debates about personal identity encounter a comparable branching problem. The Stanford Encyclopedia of Philosophy describes psychological-continuity accounts that treat causal links among memories, beliefs and other psychological states as important to persistence across time. Yet if one earlier state produces two psychologically continuous successors, a fission problem arises. One object cannot be numerically identical with two later objects that are distinct from each other. This thought experiment does not directly settle the conditions for machine subjecthood, but it offers a clear warning: similarity, inheritance and continuity are not the same relation, and copying does not automatically preserve a unique subject.
The March 2026 preprint The Artificial Self argues that machines can be copied, edited and simulated, so several identity boundaries—including instance, model and persona—may become coherent. Its authors also report that changing the identity boundary suggested to a model can alter its behaviour. The evidential limit matters. This is a preprint, and its experiments support the claim that self-reports and behaviour are sensitive to identity framing. They do not demonstrate subjective experience or establish which boundary constitutes a genuine subject. The paper's value here is that it turns the question “which AI are we counting?” from a linguistic habit into an analysable problem.
In Sustenesis Theory, identity is not a property automatically conferred by attaching a name to an object. Difference first appears as distinguishable operation after copying: different inputs, memories, tools, deployment rules and feedback place each instance in different relations. Constraint consists of the conditions that permit or prevent these differences from developing, including base weights, fine-tuning data, system prompts, computational environments, permissions and institutions. Sustained Coherence is not parameter similarity by itself. It is structural consistency maintained through constraints, feedback, correction and effective operation. If two instances do not share memory updates, feedback loops, operational boundaries and state correction, they share a technical starting point at most. Common weights alone do not form one subject persisting across instances.
My judgement is that “the same model” in current technical language normally denotes a shared lineage or compatible version, not a single subject. One model name can cover base weights, quantised variants, fine-tuned descendants, cloud services and countless temporary instances. The name is useful in engineering because it identifies provenance and approximate capabilities. It is too broad, however, when the discussion concerns knowledge, memory, intention or responsibility. To say that “the model knows” something should at least specify the version, deployment, external memory and whether the information can persist, be retrieved, tested and revised across sessions. Without those details, a local structure formed in one run may be mistakenly attributed as a permanent property of the entire model family.
This judgement does not imply that a machine could never become a subject. It says only that open weights, parameter equality and behavioural similarity are insufficient evidence of subject identity. A stronger conclusion would require evidence that a particular system has causally continuous autobiographical memory; that it continues to use those memories as its own history; that it maintains a distinguishable self–environment boundary through change and copying; and that one operating structure carries its goals, corrections and normative judgements over time. If such conditions emerge, subjecthood should be assessed at the level of that concrete, persisting system rather than substituted by a model name. The facts currently published about Hy4 preview establish the formation of a broad technical lineage. They do not establish that the lineage itself is a subject.
References
Tencent, Tencent Releases and Open-Sources Tencent Hy4 preview, 28 August 2026
https://www.tencent.com/tencent-releases-and-open-sources-tencent-hy4-preview/
Tencent, Hy4 preview FP8 repository and fine-tuning files, 28 August 2026
https://huggingface.co/tencent/Hy4-preview-FP8/commit/9a43a99d1410b0d255a27f14ba67de6a30021ade
Reuters, China's Tencent releases new open-source AI model for coding, research tasks, 28 August 2026
https://www.reuters.com/world/asia-pacific/chinas-tencent-releases-new-open-source-ai-model-coding-research-tasks-2026-08-28/
Douglas et al., The Artificial Self: Characterising the landscape of AI identity, preprint, 11 March 2026
https://arxiv.org/abs/2603.11353
Stanford Encyclopedia of Philosophy, Personal Identity, revised 30 June 2023
https://plato.stanford.edu/entries/identity-personal/
Discover more from Geoffrey Chen
Subscribe to get the latest posts sent to your email.