News analysis · Published

Tencent Hy4 Preview: What Open-Weight AI Buyers Should Test

By the ELYMENT AI editorial team · Free to read

Tencent released Hy4 preview on 28 August 2026 under the Apache 2.0 licence. The mixture-of-experts model has 770 billion total parameters but activates about 49 billion per token, and Tencent positions it for coding, office analysis, game development and scientific research. The business lesson is that open weights can improve control and portability without making deployment simple or inexpensive. Buyers should evaluate licence rights, serving infrastructure, workflow quality, operating cost and exit options before treating Hy4 or any large open model as a practical alternative to a managed API.

A vast modular AI model core inside a dark data-centre stack beneath the headline Open Weights Can Still Be Heavy.
Original ELYMENT.AI editorial illustration.

What Tencent released

Tencent published Hy4 preview through Hugging Face and other model repositories on 28 August 2026. Its model card describes a 770-billion-parameter mixture-of-experts architecture with roughly 49 billion parameters activated per token, 78 layers and a context window of up to one million tokens. The weights are available under Apache 2.0, while reference serving instructions use the vLLM and SGLang frameworks.

Reuters reported that Tencent is targeting software engineering, research and financial-analysis work. Tencent's own materials broaden that positioning to coding, office analysis, game development and scientific research. It is a preview release: the model card warns that Hy4 may reason longer than necessary and over-verify answers in some cases. Those caveats matter when latency and compute spend are part of the business case.

Open weights do not remove infrastructure

A permissive licence can reduce some legal and platform restrictions, but it does not turn a very large model into a laptop deployment. Tencent's reference recipes for the FP8 version use eight-way tensor parallelism. Hugging Face also states that no hosted inference provider currently serves the model from its platform. A business wanting direct control may therefore need specialist accelerators, distributed serving, monitoring and capacity engineering, or a separate hosting partner.

The mixture-of-experts design reduces the parameters used for each token relative to the full model, but activated parameter count is not the same as total memory, throughput or cost. The full weights still need to be stored and served. Teams should benchmark end-to-end latency, concurrency, output length, energy and operational support in the environment they expect to use, not infer economics from one architecture number.

Treat the benchmark as evidence to reproduce

Tencent reports a company-run blind evaluation involving 163 internal experts and 203 engineering tasks. Hy4 averaged 2.99 on Tencent's five-point scale, compared with 2.92 for GLM-5.3 and 2.94 for Kimi K3. Tencent also reports that Hy4 won or tied 59.6% of comparisons with GLM-5.3 and 59.1% with Kimi K3. These are useful vendor disclosures, not independent proof that Hy4 will outperform those models in another organisation's work.

Business evaluation should reproduce the decision that matters: whether the system completes a defined workflow to an acceptable standard at a sustainable cost. Use representative files, tools, failure cases and reviewer rubrics. Record unsupported claims, unnecessary tool use, excessive reasoning, recovery performance and the proportion of outputs accepted without rework.

A five-part open-model buying test

Before approving a large open-weight model, connect the licence decision to the operating model.

  • Rights: confirm commercial use, redistribution, modification and downstream obligations for the exact weights and supporting software.
  • Footprint: price the accelerators, storage, network, serving framework, engineering and support required for target concurrency.
  • Evidence: test representative workflows against managed and open alternatives using the same data, tools and acceptance rubric.
  • Controls: define data handling, access, monitoring, patching, incident response and human approval outside the model itself.
  • Portability: preserve prompts, evaluations, logs and interfaces so the business can change models or hosting environments.

What business leaders should do next

Choose one high-value workflow and compare Hy4 with the strongest managed and open alternatives available to your organisation. Require an infrastructure bill of materials, a measured cost per accepted outcome and a documented fallback. Keep the licence review separate from the performance decision: permission to run a model does not establish that it is economical, reliable or supportable.

ELYMENT AI's Kimi K3 analysis explains why model access and cloud distribution are separate commercial layers. Our GLM-5.3 review shows how staged releases affect evaluation, while the NVIDIA capacity-planning guide connects demand growth to infrastructure exposure. ELYMENT AI can help turn those considerations into a model-selection scorecard and a deployment plan that remains portable as the market changes.

Sources

Continue learning

Frequently asked questions

Is Tencent Hy4 preview open source?

Tencent has released the Hy4 preview model weights under the Apache 2.0 licence. Buyers should still review the exact weights, code, dependencies and data disclosures needed for their intended use.

How large is Tencent Hy4 preview?

Tencent's model card lists 770 billion total parameters, with roughly 49 billion activated per token, 78 layers and a context window of up to one million tokens.

What should a business test before deploying Hy4?

Test licence rights, infrastructure footprint, workflow quality, fully loaded cost, data and security controls, preview limitations and the ability to change models or hosting providers.

Explore ELYMENT AI