Tencent Hy4 Preview Review: Is the 1M-Context Coding Model Worth a Pilot?

Yes, Tencent Hy4 preview is worth a controlled coding-agent pilot if your work is limited by long context, multi-step tool use, or repeated repository retrieval. It is not yet a safe default for every production workflow. The model is enormous, self-hosting is a datacenter project, and Tencent explicitly labels this an early version that can reason and verify longer than necessary.

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.

This review was checked on 1 September 2026. It owns one narrow search intent: whether a software team should pilot Hy4 preview for long-horizon coding work. For a market-wide model shortlist, use our separate 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

What 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.

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?

Most teams should start with WorkBuddy, CodeBuddy, TokenHub, OpenRouter, or another reviewed hosted route. Tencent's published API price is $0.834 per million input tokens, $2.501 per million output tokens, and $0.042 per million cache-hit tokens. Confirm the live rate, region, retention, throughput, context limit, and tool support with the route you actually buy.

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.

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

1. Freeze 20 to 30 tasks. Include bug fixes, repository navigation, feature work, test repair, long documents, and tool failures from real codebases.
2. Keep the harness comparable. Use the same repository commits, instructions, tools, time limits, and acceptance tests for Hy4 and the current model.
3. Test context in bands. Run compact context, curated large context, and a near-window stress case. Do not assume more input is better.
4. Capture the complete cost. Record uncached input, cache hits, output, retries, elapsed time, and active human review minutes.
5. Exercise tools and recovery. Inject failed commands, malformed output, a stale file, and an unavailable dependency. Verify that the agent stops or recovers predictably.
6. Review security boundaries. Classify source code and data, restrict credentials and network access, redact logs, and require approval for production or destructive actions.
7. 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.

Recommendation

Pilot Hy4 preview when long-running engineering work is a measured bottleneck and you can compare it against a real baseline. Use hosted access first. Keep repository context structured, instrument cache behavior, and make tests rather than the model's self-confidence the acceptance gate.

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 1 September 2026 review date.