Tencent Hy4 Preview Guide 2026: 770B MoE & 1M Context
On August 29, 2026, we reviewed Hy4 preview, released by Tencent one day earlier. The important change is not merely the move from Hy3 to Hy4: Tencent scaled the backbone to 770B parameters, expanded context to 1M, and continued evaluating the model inside CodeBuddy, WorkBuddy, and real engineering workflows.
In one sentence: Hy4 preview is Tencent Hunyuan's new open MoE flagship for coding, agents, office analysis, game development, and scientific research; its official self-hosted served model name is hy4-preview, but model IDs, prices, and availability differ by platform.
This guide relies on Tencent's release, official repository, Tencent Cloud TokenHub pages, and primary papers cited by the repository. Vendor benchmarks are not independent tests by this site. At verification time, this site's live catalog did not list hy4-preview, so this article does not claim that the gateway route is already open.
Hy4 preview key facts
| Item | Verified official information |
|---|---|
| Release date | August 28, 2026 |
| Official name | Tencent Hy4 preview |
| Architecture | Mixture-of-Experts (MoE) |
| Backbone parameters | 770B total, 49B active |
| Native MTP layer | 10B total, 0.7B active |
| Backbone depth | 78 layers |
| Expert layout | 256 routed + 1 shared expert; top-8 routed experts activated per token |
| Attention / residual | Gated DSA + IndexCache; four iHC residual streams |
| Context | 1M in the repository; Tencent Cloud lists 960K maximum input and 64K maximum output |
| Official local served model | hy4-preview |
| Open weights | tencent/Hy4-preview, tencent/Hy4-preview-FP8 |
| License | Apache 2.0 |
| Tencent Cloud Guangzhou list price | CNY 6 input, CNY 18 output, CNY 0.3 cache read per 1M tokens |
Last verified August 29, 2026. Recheck the TokenHub console, target region, current price, and response usage fields before production adoption.
What is Hy4 preview, and is it the final Hy4 release?
Hy4 preview is an early Hy4 iteration, not a future final release without the preview suffix. Tencent explicitly documents known issues, including spending longer than necessary reasoning through complex tasks and over-verifying its own work.
That boundary matters. “Flagship preview” means the model already has open weights, inference recipes, a fine-tuning pipeline, and API access. It does not mean behavior, price, or service levels are permanently frozen.
Tencent also makes Hy4 preview available in WorkBuddy, CodeBuddy, Yuanbao, and ima, with API access through Tencent Cloud TokenHub and OpenRouter. That is enough to justify evaluation, while critical workflows should still pin versions, maintain regression suites, and retain a fallback model.
What changed in the 770B MoE architecture?
The Hy4 preview backbone has 78 layers. The first uses a dense FFN; the remaining 77 are MoE layers with 256 routed experts and one shared expert. Every token activates the top eight routed experts plus the shared expert.
A native MTP layer supports speculative decoding. Tencent reports the backbone as 770B total and 49B active parameters; including MTP adds 10B total and 0.7B active parameters. When one page says 770B and another approaches 780B, check whether MTP is included.
Attention uses Gated DeepSeek Sparse Attention (Gated DSA) with IndexCache for cross-layer sparse-index reuse. The residual path uses four identity Hyper-Connection (iHC) streams. Together, these designs target manageable compute and information flow at 1M context and large MoE scale.
The repository also specifies 64 attention heads, a 2,048 query-compression dimension, a 512 key-value compression dimension, and an indexer top-k of 2,048. These values help with inference-framework compatibility and capacity planning; they do not prove any task outcome by themselves.
How should you interpret the 1M context window?
Tencent's repository lists a 1M context length. The Tencent Cloud product page describes hosted limits of 960K maximum input and 64K maximum output. These can coexist because a model window and a hosted service's reserved output, system prompt, and safety margin use different accounting boundaries.
Do not start production testing with a request near one million tokens. Use tiers such as 32K, 128K, 256K, 512K, and ultra-long context, recording:
- time to first token and end-to-end latency
- input, output, and cache-read tokens
- cross-file retrieval accuracy
- tool-call success and retry count
- task pass rate and human rework time
Context capacity is not automatic correctness. Large repositories, financial workbooks, and research corpora still need retrieval, permission isolation, and evidence links.
How strong is Hy4 preview for coding and agents?
Tencent's benchmark appendix reports broad gains over Hy3 in software engineering, tool use, and long-context tasks. Selected values from the same official table are below:
| Tencent-published evaluation | Hy3 | Hy4 preview |
|---|---|---|
| SWE-bench Multilingual | 75.8 | 82.9 |
| DeepSWE | 28.0 | 64.3 |
| SWE Atlas - Refactoring | 32.9 | 53.3 |
| Terminal-Bench 2.1 | 70.8 | 85.4 |
| Toolathlon-Verified | 56.2 | 74.1 |
| OneMillionBench (with tools) | 51.5 | 65.4 |
These are Tencent-published results, not independent tests by this site. Tencent notes that models were evaluated at their highest available reasoning setting, some starred results were run by Tencent, and harnesses, token budgets, sandbox resources, and repeated sampling affect scores.
The same table prevents cherry-picking: Hy4 preview scores 17.5 on ProgramBench, below Kimi K3, GPT 5.6 Sol, and Claude Opus 5 in Tencent's comparison. The generation-over-generation gain is real, but Hy4 does not lead every coding benchmark.
What does Tencent's 203-task expert evaluation show?
Tencent ran a blind evaluation with 163 internal experts across 203 real engineering tasks. The published result was:
- Hy4 preview: 2.99 / 4 average
- GLM 5.3: 2.92 / 4; Hy4 won 46.8%, tied 12.8%, and lost 40.4%
- Kimi K3: 2.94 / 4; Hy4 won 51.2%, tied 7.9%, and lost 40.9%
This matters because the tasks came from Tencent's software engineering, gaming, finance, and security work rather than only public leaderboards. It remains an internal vendor evaluation, not a fully reproducible third-party benchmark with a public task set and rubric.
The reliable adoption method is to build a regression suite from your own repositories, documents, and acceptance commands, then compare pass rate, latency, and total cost against the models you already run.
What workloads fit Hy4 preview?
- long-horizon coding agents that plan, patch, debug, and verify
- frontend work where code, visual hierarchy, and interaction quality all matter
- cross-file office workflows that deliver documents, spreadsheets, and presentations
- playable game prototypes followed by multi-turn engine work
- financial modeling and multi-file data analysis
- research in AI, molecular dynamics, condensed-matter physics, and mathematics
- tool-using search and knowledge workflows over 1M context
Production write actions still need policy gates or human approval for shell, databases, payments, permissions, and external messages. A stronger model does not remove agent authorization risk.
What is the official Hy4 preview price?
Tencent Cloud TokenHub currently shows these reference prices for Guangzhou:
| Billing item | Public price |
|---|---|
| Input | CNY 6 / 1M tokens |
| Output | CNY 18 / 1M tokens |
| Cache read | CNY 0.3 / 1M tokens |
This is a Tencent Cloud region- and time-specific direct price, not a universal price for third-party platforms. OpenRouter, other inference providers, and self-hosting use separate economics; a gateway multiplier cannot be inferred from TokenHub's direct price.
At verification time on August 29, 2026, this site's live pricing page did not list hy4-preview. If it is added later, use the live catalog and request ledger rather than the Tencent Cloud direct price in this article.
Do not mix these Hy4 model IDs
| Access path | Name / ID to use |
|---|---|
| Official repository's local served model | hy4-preview |
| OpenRouter | tencent/hy4-preview |
| Hugging Face BF16 weights | tencent/Hy4-preview |
| Hugging Face FP8 weights | tencent/Hy4-preview-FP8 |
An API gateway usually does not translate a Hugging Face repository name into a hosted model ID. Copy the exact identifier from the target platform and verify it with a model-list call or a minimal request.
How do you self-host and call Hy4 preview?
Tencent recommends vLLM or SGLang. After launch, call the local served model through an OpenAI-compatible Chat Completions API:
from openai import OpenAI
client = OpenAI(
base_url="http://127.0.0.1:8000/v1",
api_key="EMPTY",
)
response = client.chat.completions.create(
model="hy4-preview",
messages=[
{"role": "user", "content": "Review this patch and list the highest-risk regressions."}
],
temperature=0.9,
top_p=1.0,
)
print(response.choices[0].message.content)
Tencent recommends temperature=0.9 and top_p=1.0. The model defaults to high reasoning. The repository's local template uses chat_template_kwargs.reasoning_effort = "no_think" for direct answers; hosted providers may expose different extension fields.
The official vLLM recipe uses eight-way tensor parallelism, MTP speculative decoding, the FLASHMLA_SPARSE attention backend, and hy_v4 tool and reasoning parsers. Do not paste the recipe onto hardware with different GPU, driver, image, or framework support without a capacity check.
Hy4 preview vs Hy3
| Dimension | Hy3 | Hy4 preview |
|---|---|---|
| Official scale | 295B total / 21B active | 770B total / 49B active |
| Context | 256K | 1M |
| Positioning | cost-efficient agents and productivity | new flagship for coding, agents, and complex productivity |
| Maturity | progressed from preview to formal release | early Hy4 preview |
| Self-hosting cost | lower | materially higher |
Hy3 remains practical when cost, deployment scale, and maturity dominate. Add Hy4 preview to a controlled evaluation when you need longer context and stronger long-horizon engineering or tool execution, but do not replace a production model based only on total parameter count.
Production evaluation checklist
- Copy the exact model ID from the target platform; do not guess casing or namespace.
- Pin the model version, prompt, tools, reasoning effort, timeout, and maximum output.
- Start with read-only tasks before granting file, shell, database, or external side effects.
- Measure pass rate with real repositories and acceptance commands, not chat impressions.
- Test context tiers from 32K upward and record latency, cache usage, and total billing.
- Set maximum turns, token budgets, and approval gates for complex tasks.
- Verify tool calls, structured output, and reasoning-field mapping on the target provider.
- Keep a regression suite, fallback route, and fast rollback for a preview model.
Key takeaways
- Hy4 preview launched and was open sourced on August 28, 2026 as Tencent Hunyuan's new MoE flagship.
- Its backbone has 770B total and 49B active parameters, plus a native 10B/0.7B MTP layer.
- The repository lists 1M context; Tencent Cloud lists 960K maximum input and 64K maximum output.
- Tencent reports major gains over Hy3 in coding, tool use, and long context, but not leadership on every benchmark.
hy4-preview,tencent/hy4-preview, andtencent/Hy4-preview-FP8are platform-specific identifiers.- This gateway had not opened the route at publication time; availability and billing require a live-catalog check.
Frequently asked questions
Has Hy4 preview officially launched?
Yes. Tencent released and open sourced Hy4 preview on August 28, 2026, with access through WorkBuddy, CodeBuddy, Yuanbao, ima, TokenHub, and OpenRouter. It remains a preview, not the final Hy4 release.
Is Hy4 preview multimodal?
Tencent's launch repository presents it as a language model for coding, agents, and productivity, with public specifications focused on text, tools, and long context. Do not infer image or video input merely because the wider Hunyuan family includes visual models; check the target provider's capability table.
What is the Hy4 preview context length?
The repository specifies 1M context. Tencent Cloud separately lists 960K maximum input and 64K maximum output for the hosted service.
What is the official Hy4 preview API ID?
The official repository's served model name is hy4-preview; OpenRouter uses tencent/hy4-preview; self-hosted weights use tencent/Hy4-preview or the FP8 variant. The target platform's console ID takes precedence.
Does Hy4 preview support tool calling?
Tencent's vLLM recipe enables the hy_v4 tool parser and automatic tool choice, while its benchmark suite includes Toolathlon and MCP-Atlas. Verify exact hosted parameters with the provider.
How much does Hy4 preview cost?
Tencent Cloud TokenHub currently shows CNY 6 input, CNY 18 output, and CNY 0.3 cache read per million tokens in Guangzhou. Region, platform, and promotions can change the price.
Can I call Hy4 preview through this gateway now?
At verification time on August 29, 2026, the live catalog and enabled upstreams did not list hy4-preview. Check the live pricing page for any later availability.
Can Hy4 preview work with Claude Code or another coding agent?
The model targets coding and agents and supports a local OpenAI-compatible serving path. Client compatibility still depends on protocol, tool-call shape, reasoning output, and the provider's compatibility layer, so validate a minimal request and tool regression first.
Primary sources
- Tencent release: Tencent Releases and Open-Sources Tencent Hy4 preview
- Official Tencent Hunyuan GitHub: Hy4-preview
- Tencent Cloud TokenHub product and pricing page
- Tencent Cloud Hunyuan product page: Hy4 preview limits
- Gated DeepSeek Sparse Attention paper
- IndexCache paper
- This site's live model pricing: for later route availability, not Tencent Cloud's official price