Tencent Hy4 preview

Tencent's open-weight flagship, released August 28, 2026. A 770B MoE with 49B active, a 1M-token context, Apache 2.0 — and explicitly a preview.

Last reviewed September 6, 2026

At a glance

Overview

Tencent Hy4 preview is Tencent’s open-weight flagship large language model, released and open-sourced on August 28, 2026. It is a Mixture-of-Experts transformer with 770 billion total parameters and 49 billion activated per token, a 1M-token context window, and Apache 2.0 weights published at tencent/Hy4-preview.

The pitch is narrower than a generic frontier-model launch. Tencent’s announcement says Hy4 preview “is built for real-world productivity tasks, delivering outstanding performance across coding, office work, and scientific research,” and describes it as “developed using high-quality training data co-created with Tencent experts across software engineering, gaming, finance, security, and other domains, as well as through deep co-design with products such as WorkBuddy.”

That co-design story is the interesting part. Hy4 preview was built alongside two shipping Tencent products — CodeBuddy, the coding assistant, and WorkBuddy, the workplace assistant — and trained on work data contributed by software engineers, game developers, finance analysts and security experts. The model is shaped by the workloads those products see, rather than aimed at a leaderboard. Whether that produces a better assistant on your work is exactly what the vendor’s own evaluation cannot tell you.

A note on naming

Tencent’s own spelling is “Tencent Hy4 preview”, lowercase preview, and the Hugging Face repository is tencent/Hy4-preview. Third-party coverage frequently writes “Hunyuan 4” or “Hy4 Preview”; those refer to the same model. Neither should be confused with Hunyuan-4B, an unrelated 4-billion-parameter edge model from July 2025.

Until August 28, 2026 there was no Hunyuan 4 flagship at all, and earlier versions of this page said so. That is no longer true — the release is real, and this page now documents it.

The previous flagship: Tencent Hy3

Tencent Hy3 has not been retired. Released July 6, 2026, it is a 295B MoE with 21B active parameters, 192 experts with top-8 activation, a single 3.8B Multi-Token Prediction layer for speculative decoding, a 256K context window and Apache 2.0 weights. Its distinguishing feature was hybrid fast-and-slow thinking via a reasoning_effort parameter with no_think (default), low and high settings.

Hy3 remains listed on Tencent Cloud TokenHub, its model card is still published, Tencent has announced no deprecation date, and free Hy3 access on WorkBuddy and CodeBuddy runs through September 30, 2026. At 21B active parameters against Hy4 preview’s 49B it is also the cheaper model to self-host by a wide margin, and for latency-sensitive work its no_think default is a genuine advantage over a model that defaults to deep reasoning.

Capabilities

• Long-horizon software engineering: Tencent claims “stronger understanding, planning, debugging, and validation capabilities for long-context development tasks” — the debugging-across-a-whole-repository case rather than single-function completion.
• Document and spreadsheet production: Enhanced “data analysis and cross-document collaboration, supporting the full workflow from information processing through to the creation of documents, spreadsheets, and presentations.” Turning messy multi-file context into a deliverable is the named target.
• Game prototyping: Tencent claims the model can “generate a playable prototype from a single natural-language request” — plausible given the game-developer training data, and unverified outside Tencent.
• Scientific reasoning: Improvements claimed in “molecular dynamics simulation, condensed-matter physics and fundamental mathematics.”
• Deep reasoning by default: The model card defaults reasoning_effort to high (deep chain-of-thought); direct answers require explicitly passing no_think. This is the opposite of Hy3’s default and it changes both latency and token spend.
• 1M-token context: A million-token context window with sparse attention, aimed at whole-repository and multi-document work rather than single prompts.

Technical Specifications

• Hugging Face repositories: tencent/Hy4-preview, tencent/Hy4-preview-FP8 (mirrors on ModelScope, GitCode, CNB)
• API model ID: hy4-preview
• Total parameters: 770B
• Active parameters: 49B per token
• Layers: 78 — one dense FFN layer, 77 MoE layers
• Experts: 256 routed + 1 shared per MoE layer; top-8 routed experts activated per token
• Attention: Gated DeepSeek Sparse Attention (Gated DSA) with IndexCache for cross-layer sparse index reuse
• Residual pathway: iHC, for expanded inter-layer information flow
• Vocabulary size: 120,832
• Context window: 1M tokens (Tencent Cloud documents 960K max input, 64K max output)
• License: Apache 2.0
• Thinking control: reasoning_effort defaults to high; override via extra_body={"chat_template_kwargs": {"reasoning_effort": "no_think"}}
• Serving: vLLM and SGLang, both with official prebuilt images

Tencent publishes no knowledge cutoff and no training-token count for Hy4 preview, so this page does not state one. Ground time-sensitive queries with retrieval or tool use.

Use Cases

• Repository-scale debugging: The 1M context plus the long-horizon planning claim targets the case where the bug is not in the file you are looking at. This is the workload CodeBuddy was co-designed around.
• Document and spreadsheet generation from scattered sources: Pulling a quarter’s worth of tickets, notes and exports into a single deliverable — the WorkBuddy workload, and the one the training-data provenance most directly supports.
• Cost-constrained hosted inference: At $0.834 / $2.501 per million tokens it undercuts every Western frontier model by an order of magnitude, which makes it viable for high-volume classification, extraction and summarisation.
• Self-hosting where data cannot leave: Apache 2.0 with no field-of-use restrictions — unlike several competing Chinese open-weight releases — but 770B total parameters means a serious cluster even at FP8.
• Game prototyping loops: A named target domain with matching training data; worth a scoped trial if that is your workload, and not worth assuming otherwise.
• Scientific and mathematical drafting: Molecular dynamics and pure maths are called out explicitly. Treat as a starting point for verification, not a source of results.

Performance / Benchmarks

Tencent publishes most of its benchmark appendix as an image rather than as text, which limits what can be quoted precisely. The scores stated in the model card text are:

These are vendor-reported figures, run by Tencent, published at launch. Comparator columns exist in the appendix image but are not reproduced here, because a number read off a chart is not a citable number.

No independent aggregate score — such as an Artificial Analysis Intelligence Index v4.2 entry — had been published for Hy4 preview as of September 6, 2026.

The blind expert evaluation, in full

Tencent’s headline comparison is a blind internal study, and every part of that phrase matters:

• Who ran it: Tencent, internally.
• Who judged: 163 Tencent experts.
• What they judged: 203 engineering tasks, selected by Tencent.
• Result: Hy4 preview averaged 2.99 / 4.00, against Kimi K3 at 2.94 and GLM-5.3 at 2.92.

A 0.05-point margin on a four-point scale, from a panel the vendor assembled, on a task set the vendor chose, is a near-tie. It is evidence that Hy4 preview is in the same class as the leading Chinese open-weight models; it is not evidence that it beats them, and it should not be reported as a ranking. Run your own comparison on your own tasks before deciding.

Limitations

• “Preview” is a real caveat, not branding. The model card states: “This is an early version of Hy4. There is real headroom left in both pre-training and post-training, and we are shipping with known issues.” There is no published GA date, no stability commitment for the served endpoint, and no statement that these weights are final. Building a product on it means accepting that both the checkpoint and the API may move.
• Named failure modes: Tencent lists “spending longer than necessary reasoning through complex tasks, and a tendency to over-verify its own work.” Both cost tokens and latency.
• Deep reasoning is the default. Unlike Hy3, reasoning_effort starts at high. A naive integration will be slower and more expensive than the price card suggests until no_think is set explicitly.
• 770B is a lot to self-host. Apache 2.0 removes the licence cost, not the hardware cost. Hy3 at 295B/21B active, or a smaller open-weight model, remains the realistic self-hosting option for most teams.
• Vendor-reported benchmarks, image-only appendix. Every published number originates with Tencent, and most of the comparison table is a picture. Independent replication was thin at the time of writing.
• The blind evaluation is vendor-run with a 0.05-point margin over its comparators. See above.
• Weights, not full openness. Apache 2.0 covers the weights; training data and pipeline are not released.
• No published knowledge cutoff.
• The free windows expire. Free Hy4 preview access ran two weeks from launch and has closed; free Hy3 access ends September 30, 2026.

Pricing & Access

Self-hosting is free of licence cost. The weights are Apache 2.0 at tencent/Hy4-preview, with an FP8 checkpoint for cheaper serving. Commercial use, modification and redistribution are permitted.

Hosted USD rate card. Tencent’s own English announcement states: “API pricing is set at USD 0.834 per million input tokens, USD 2.501 per million output tokens and USD 0.042 per million tokens for cache hits.” Tencent Cloud documents the same figures for its Singapore endpoint. These are Tencent’s published USD prices, not a conversion made by this page.

Domestic CNY rate card. Tencent Cloud TokenHub’s 模型价格 (“Model Pricing”) page, last updated 2026-09-04, prices the same models in 元 per million tokens:

All figures in that table are yuan (CNY), unconverted. For what it is worth, the two cards imply a rate near 7.2 CNY to the dollar across all three columns — but the USD figures above are the ones Tencent publishes for international customers, and they are what the frontmatter quotes.

Endpoints: Singapore at https://tokenhub-intl.tencentcloudmaas.com/v1, Silicon Valley at https://tokenhub-us.tencentcloudmaas.com/v1. The model is also routed through OpenRouter as tencent/hy4-preview at the same USD prices.

Free access, and its dates. Tencent offered free Hy4 preview access on WorkBuddy and CodeBuddy for two weeks from the August 28, 2026 launch — that window has closed. Free Hy3 access on the same two products continues through September 30, 2026.

Consumer access runs through Yuanbao, ima and other Tencent applications.

Ecosystem & Tools

• tencent/Hy4-preview on Hugging Face — Apache 2.0 weights, FP8 checkpoint, deployment guides
• tencent/Hy3 on Hugging Face — the previous flagship, still published
• Hunyuan portal — Tencent’s model hub
• Tencent Cloud TokenHub 模型价格 — the CNY rate card for Hy4 preview and Hy3
• OpenRouter — third-party routing at tencent/hy4-preview
• CodeBuddy / WorkBuddy — the coding and workplace assistants the model was co-designed with
• Yuanbao, ima — consumer assistants
• vLLM and SGLang — supported serving paths, both with official prebuilt images

Community & Resources

• Tencent Releases and Open-Sources Tencent Hy4 preview — the official announcement, dated August 28, 2026
• Hy4 preview model card — architecture, known issues, deployment
• Hy4 preview FAQ — Tencent Cloud on regions, limits and context budgeting
• Tencent AI blog
• Compare with GLM-5.3, DeepSeek V4, Kimi K3, Qwen3.8-Max, Ling-2.6-1T and LongCat-2.0