In this piece
Tencent Hy4 Preview Review: What Changes with the Full Hy4 Release?
For a new hosted pilot, start with Hy4, the successor Tencent released on 4 November 2026. It adds image understanding and Tencent reports improvements on long-running agent tasks. Our earlier WorkBuddy test, the specifications below and the self-hosting recipe refer to the August Preview. They do not establish the full release’s benchmark results, hardware requirements or licensing.
The Tencent Cloud retirement notice sets 7 December 2026 at 00:00 Beijing time as the cutoff for hy4-preview on TokenHub, Agent Development Platform and CloudBase. Test hy4 before the cutoff. Eligible users may be switched automatically; other calls can fail. Check model identifiers, tool calls, structured output, budgets and fallbacks on your provider before changing production.
Tencent released Hy4 preview on 28 August 2026 with 770 billion backbone parameters, 49 billion active parameters per token, and a context window above one million tokens. The official Tencent announcement also lists WorkBuddy, CodeBuddy, Tencent Cloud TokenHub, and OpenRouter as access routes. Launch access in WorkBuddy and CodeBuddy was announced as free for two weeks.
We rechecked the rollout and pricing on 11 November 2026. For a market-wide shortlist, see our open-weight LLM comparison. For the architecture choice between retrieval, tuning, and large context, use the RAG vs fine-tuning vs long-context guide.
Hy4 preview facts in 60 seconds
| Question | Verified answer | Buyer implication |
|---|---|---|
| What is it? | A 770B-total, 49B-active Mixture-of-Experts language model | Large capacity with sparse activation, not a 49B checkpoint |
| How much context? | 1,048,576 tokens in the serving recipe | Enough capacity for very large inputs, but capacity is not proof of recall |
| Is it open? | Weights and code under Apache 2.0 | Commercial use and self-hosting are possible, subject to a normal legal review |
| Can it call tools? | Yes, official vLLM and SGLang recipes include reasoning and tool parsers | Agent integration is supported, but your harness still owns permissions and recovery |
| August Preview list price | $0.834 input, $2.501 output, and $0.042 cache hit per 1M tokens | Long prompts can be affordable when cache reuse is real |
| Can a team self-host it? | Yes, with official recipes targeting 8 B300 or 16 B200 GPUs | Hosted evaluation should come before infrastructure procurement |
| Main known limitation? | Overthinking and over-verification on complex work | Measure time, output volume, and cost per accepted change |
Need a neutral Hy4 pilot with real repository tasks, cost controls, and acceptance gates?
Plan the Two-Week PilotWhat is Tencent Hy4 preview?
Hy4 preview is Tencent Hy Team’s flagship sparse Mixture-of-Experts model for coding, office work, analysis, games, and scientific tasks. The official Hugging Face model card documents 78 backbone layers. The first uses a dense feed-forward network; the other 77 contain 256 routed experts and one shared expert. Each token activates eight routed experts plus the shared expert.
The 770B figure covers the backbone. A native multi-token-prediction layer adds another 10B total parameters and 0.7B active parameters for speculative decoding, which explains why some model listings show 780B. The architecture combines Gated DeepSeek Sparse Attention, cross-layer IndexCache reuse, and identity Hyper-Connections. The practical goal is to make long inputs and sustained reasoning more efficient without activating the full model for every token.
The weights, an FP8 variant, fine-tuning code, deployment guidance, and license are available in the official Tencent Hy4 preview repository. That is meaningful openness. It does not make a 770B model a laptop download.
What did our WorkBuddy coding test show?
We asked Hy4 preview inside WorkBuddy to build a fully interactive AI model benchmark dashboard. The useful result was not that it produced HTML. It also wrote the core logic, built headless harnesses, verified JavaScript against a stub DOM, and checked filters plus cost-simulation edge cases before delivery.
That is promising evidence of workflow completion, not an independent benchmark. One successful dashboard does not establish a general success rate, security level, or economic advantage. It does reveal a good pilot hypothesis: Hy4 may be valuable when an agent must retain requirements, implementation state, test evidence, and tool output across a longer run.
How strong are the Hy4 preview benchmarks?
Tencent reports that 163 internal experts blindly rated outputs on 203 engineering tasks. Hy4 averaged 2.99 out of 4, narrowly above GLM-5.3 at 2.92 and Kimi K3 at 2.94. Against GLM-5.3, Tencent reports 46.8% wins, 12.8% ties, and 40.4% losses. Against Kimi K3, the result was 51.2% wins, 7.9% ties, and 40.9% losses.
| Signal | Published result | What it supports | What it does not prove |
|---|---|---|---|
| Tencent expert evaluation | 2.99/4 across 203 engineering tasks | Competitive performance on Tencent’s real-work task set | Independent superiority or your repository success rate |
| SWE-bench Pro record | 65.7 | Strong issue-resolution capability under that harness | Production maintainability, latency, or safety |
| DeepSWE record | 64.3 | Useful coding-agent potential | Reliable million-token codebase understanding |
| SWE-bench Multilingual record | 82.9 | Promising multilingual code work | Equal performance across your languages and frameworks |
Use the results to justify an evaluation, not a migration. Vendor and public benchmark runs differ in harnesses, tools, context, compute, retry policy, and scoring. Hy4 is also a preview, so providers can change serving behavior while the model remains under active development.
Does a 1M-token window mean the model remembers a whole codebase?
No. A one-million-token context window is an input capacity, not a memory or accuracy guarantee. It can reduce destructive chunking and let the model inspect broader relationships in one run. It does not guarantee that a detail near the middle will influence the correct edit, that generated code will stay consistent, or that every token is worth paying to send.
The LongCodeBench research paper found that long-context coding performance can fall sharply as input grows, even when a model technically supports the window. Hy4 was released after that study and was not evaluated in it. The study supports the testing method, not a claim about Hy4's score.
For production work, provide a repository map, task boundaries, changed-file hints, acceptance tests, and concise prior evidence. Our coding-agent context guide explains why structured context usually beats an undifferentiated repository dump.
Hosted API or self-hosted Hy4?
For the full release, the TokenHub rate card lists USD 0.834 input, 2.501 output and 0.042 cache-hit input per million tokens in its Singapore table, with a 50% Hy4 inference discount through 31 December 2026 Beijing time. Budget at list prices and verify the region, endpoint and applicable terms. Keep the August Preview hardware and pricing evidence separate.
The official vLLM recipe lists the full deployment at 16 NVIDIA B200 or 8 B300 GPUs. Official vLLM and SGLang examples expose an OpenAI-compatible endpoint, use the FP8 model, and enable dedicated Hy4 reasoning and tool-call parsers. This is viable infrastructure for a team that already operates distributed inference. It is not a sensible first experiment for a normal product team.
| Route | Choose it when | Main risk | Decision gate |
|---|---|---|---|
| WorkBuddy or CodeBuddy | You want the fastest end-to-end product trial | Less control over harness details and telemetry | Can it complete your real workflow with inspectable evidence? |
| Hosted API | You need integration, measured usage, and tool calls | Provider-specific limits, data terms, and serving variance | Does cost per accepted task beat the current route? |
| Self-hosted FP8 | You have high stable volume, data-control needs, and distributed-GPU operations | Capital, utilization, networking, upgrades, and on-call ownership | Does measured TCO beat a contracted API at required reliability? |
Use our local model versus API break-even model before turning open weights into a hardware purchase. The right denominator is accepted tasks, not theoretical tokens.
A two-week Hy4 coding-agent pilot
- Freeze 20 to 30 tasks. Include bug fixes, repository navigation, feature work, test repair, long documents, and tool failures from real codebases.
- Keep the harness comparable. Use the same repository commits, instructions, tools, time limits, and acceptance tests for Hy4 and the current model.
- Test context in bands. Run compact context, curated large context, and a near-window stress case. Do not assume more input is better.
- Capture the complete cost. Record uncached input, cache hits, output, retries, elapsed time, and active human review minutes.
- Exercise tools and recovery. Inject failed commands, malformed output, a stale file, and an unavailable dependency. Verify that the agent stops or recovers predictably.
- Review security boundaries. Classify source code and data, restrict credentials and network access, redact logs, and require approval for production or destructive actions.
- Decide from accepted outcomes. Expand only if Hy4 improves first-pass acceptance, cost per accepted change, review time, or a strategic data-control requirement.
What should a commercial evaluation measure?
- First-pass acceptance rate: tasks accepted without a second implementation attempt.
- Cost per accepted change: model cost, retries, infrastructure, and review divided by accepted changes.
- Long-context retrieval accuracy: evidence found and used correctly at several input positions.
- Tool completion rate: valid calls, correct arguments, error recovery, and clean stop behavior.
- Verification quality: relevant tests run, failures interpreted correctly, and unsupported claims avoided.
- P50 and P95 lead time: the preview’s overthinking may hide inside tail latency.
- Security exceptions: denied actions, secret exposure, unapproved network access, and manual overrides.
If one model cannot win every lane, keep the routing decision separate from this review. Our multi-model coding-agent buying guide shows how to assign planning, implementation, and independent verification without hardcoding one vendor into the workflow.
Production AI help
Building an AI product and worried about inference cost, architecture, or production readiness? Wavect helps founders turn AI prototypes into reliable production systems.
Explore the service path:
Recommendation
Pilot the full hosted Hy4 when long-running engineering work is a measured bottleneck. Replay the same tasks with the same tools, budget and acceptance tests before changing the default. Keep the old Preview measurements separate.
Wait if your tasks fit comfortably inside a smaller context, your codebase lacks reliable tests, or nobody owns provider risk and agent permissions. Self-host only when a validated workload, data-control requirement, and stable utilization justify 8 to 16 datacenter GPUs plus the people required to operate them.
Primary sources and freshness boundary
Architecture, access routes, pricing, the 203-task internal evaluation, and the two-week launch offer come from Tencent’s 28 August announcement. Detailed specifications, public evaluation records, limitations, and Apache 2.0 licensing come from the Tencent model card and repository. Hardware guidance comes from the vLLM recipe. LongCodeBench provides a general method for questioning million-token coding claims; it did not test Hy4. Wavect did not reproduce Tencent’s benchmark suite. Product access, provider prices, privacy terms, and serving limits can change after the 11 November 2026 review date.
Frequently Asked Questions About Tencent Hy4 Preview
Should I start a new pilot on Hy4 Preview?
What is Tencent Hy4 preview?
Does Hy4 preview really support one million tokens?
Is Hy4 preview open source?
How much does the Hy4 preview API cost?
Can Hy4 preview run on a laptop or one GPU?
Is Hy4 preview better than GLM-5.3 or Kimi K3?
What is the biggest Hy4 preview risk for coding agents?
Final thoughts
Pilot the full hosted Hy4 when long-running engineering work is a measured bottleneck. Replay the same tasks with the same tools, budget and acceptance tests before changing the default. Keep the old Preview measurements separate.
