# Tencent Hunyuan Hy4 Preview Local Deployment and Cost-Effectiveness

> What it takes to self-host Tencent Hunyuan Hy4 preview: official BF16 and FP8 weight sizes, B200/B300/H200 configurations, 200-person and public-API tiers, Apache-2.0 duties, and API break-even math.

- Written by: ChinaAPI Research
- Published: 2026-09-07 (last updated 2026-09-07)
- Canonical: https://chinaapi.ai/insights/hy4-preview-local-deployment-value/
- Try the models: https://dash.chinaapi.ai/register?lang=en&utm_source=chinaapi&utm_medium=md&utm_campaign=hy4-preview-local-deployment

The short answer: **Hy4 preview is genuinely open-weight and commercially usable under Apache-2.0, but it is a datacenter model—not a local-workstation model.** Its 49B activated backbone parameters describe compute per token. They do not shrink the 770B backbone, the native 10B MTP layer or the roughly 758 GiB official FP8 weight files that must be stored and served.

For most teams, including a 200-person company, start with the managed API and buy a Blackwell deployment only after measured, sustained traffic or a hard data-boundary requirement justifies it. That is our recommendation, not a claim that self-hosting has no value: Hy4 is a strong fit for private long-horizon coding, office analysis, game-development and scientific-agent workloads when infrastructure control matters.

This is a deployment and cost analysis, not a ChinaAPI quality benchmark. ChinaAPI has not reproduced the open checkpoint on its own hardware, and Tencent's published quality evaluations remain vendor results.

## What actually has to fit

Tencent specifies a 770B-parameter MoE backbone with 49B backbone parameters activated per token, plus one native MTP layer with 10B total and 0.7B activated parameters. The advertised maximum context is 1,048,576 tokens. The model is text-only.

We read the public Hugging Face repository metadata on 2026-09-07 and summed the safetensor file sizes. The result is about 1,559.98 GB (1,452.85 GiB) for the original checkpoint and 813.77 GB (757.88 GiB) for the official MXFP8 checkpoint. Those are weight files before runtime allocations and KV cache.

| Official path | Published shape | Configured context in the framework matrix | What it establishes |
|---|---:|---:|---|
| MXFP8 on B300 or GB300 | 4 GPUs, TP4, one node | 262K | Smallest published GPU-count recipe |
| MXFP8 on B200 | 8 GPUs, TP8, one node | 262K | Practical single-node Blackwell baseline |
| BF16 on B300 | 8 GPUs, TP8, one node | 262K | One-node original-precision path |
| BF16 on H200 | 16 GPUs, TP16, two nodes | 131K | Hopper path; MXFP8 requires Blackwell |

The important context detail is easy to miss. SGLang estimates about 95 KB of replicated KV and sparse-index cache per token per tensor-parallel rank, then recommends 262K context on the 192GB-or-larger Blackwell paths and 131K on H200. **One million positions in the model specification is not evidence that one million tokens will meet a production concurrency or latency target on these launch shapes.**

## Three deployment tiers

### 1. Smallest official configuration

The smallest published GPU count is **4x B300 or 4x GB300 with the official MXFP8 checkpoint, TP4 and 262K configured context** in the SGLang Hy4 matrix. Treat this as the smallest documented framework path, not an absolute hardware minimum across all future quantizations and engines.

Tencent's top-level vLLM and SGLang examples use TP8. vLLM supplies a Hy4-specific image, MTP speculative decoding, the `FLASHMLA_SPARSE` attention backend and Hy4 reasoning/tool parsers. SGLang also publishes hardware-specific cells and says unlisted parallelism combinations are experimental until run end to end.

There is no credible official single-workstation route today. Even the FP8 safetensors alone are roughly six times a 128GB unified-memory workstation, before cache and runtime headroom.

### 2. A 200-person internal team

For our standard planning profile—80–120 daily active users, 20–30 peak online and 8–16 concurrent generations—start load testing with **one 8x B200 or 4x B300 FP8 worker**. Use real repository sizes, prompt lengths, tool loops, reasoning settings and output lengths. Tencent and the serving frameworks do not publish a Hy4 throughput number for this profile, so employee count cannot be converted directly into GPU count.

For a noncritical pilot, one worker is enough to learn. For an internal service that must survive maintenance or a node failure, the deployable configuration is **two independent workers**. Prefix caching, request queues and a short-context lane can materially change capacity; a 1M-context request class should not share an unrestricted queue with ordinary coding traffic.

Our stance: **do not buy two nodes for 200 employees merely because Hy4 is open-weight.** Prove that the managed route passes the task suite, then measure whether traffic is steady enough, sensitive enough or expensive enough to justify dedicated capacity.

### 3. A public commercial API

Start with at least two independently deployable workers, admission control, per-tenant limits, rolling-update capacity and separate service objectives for short and long context. Decide whether low-latency traffic should use MTP: SGLang recommends it for interactive replies but notes that draft-and-verify overhead can lose at saturation, where MTP-off high-throughput serving may be preferable.

The official recipes expose an OpenAI-compatible endpoint and parse Hy4 reasoning plus tool calls. That does not create an SLA by itself. Validate tool-schema behavior, timeouts, cache isolation, failure recovery and the exact preview-model retirement policy of any upstream route before promising stability to customers.



## API versus renting an 8x B200 node

ChinaAPI currently lists `hy4-preview` at $0.8889 per million input tokens and $2.6667 per million output tokens. The rate is generated from the gateway-reconciled catalog; the live pricing page remains authoritative.

As a reproducible infrastructure baseline, Lambda listed an 8x B200 instance at USD 6.69 per GPU-hour when reviewed on 2026-09-07, or **USD 53.52 per node-hour**, before tax. Storage, deployment labor, idle capacity, monitoring and redundancy are excluded. We have not established that ChinaAPI's managed route is byte-identical to either downloadable checkpoint, so this is an operating-cost threshold, not a like-for-like quality result.

To spend USD 53.52/hour through the API, the workload would need to sustain:

- **20.07 million output tokens/hour (5,575 output tokens/second)** if every billed token is output.
- **40.14 million total tokens/hour at a 3:1 input-to-output mix (11,150 total tokens/second, including 2,787 output tokens/second)** at three input tokens per output token.

Those thresholds are deliberately demanding. A rented node that generates below the break-even rate is more expensive on raw hourly cost before operations; a second redundancy node doubles the hurdle. Self-hosting can still win for privacy, offline operation, serving research, custom quantization or contractual control, but low utilization does not become economical just because the weights are free.

You can reproduce both the weight-size and price calculation:

```python
import json
import urllib.request

def safetensor_gib(repo):
    url = f"https://huggingface.co/api/models/{repo}?blobs=true"
    data = json.load(urllib.request.urlopen(url))
    total = sum(
        item.get("size") or item.get("lfs", {}).get("size") or 0
        for item in data["siblings"]
        if item["rfilename"].endswith(".safetensors")
    )
    return total / 2**30

print(safetensor_gib("tencent/Hy4-preview"))
print(safetensor_gib("tencent/Hy4-preview-FP8"))

# Download /data/model-pricing.json from this site beside the script.
rows = json.load(open("model-pricing.json", encoding="utf-8"))["models"]
row = next(item for item in rows if item["model"] == "hy4-preview")
node_hour = 8 * 6.69
output_tokens_per_hour = node_hour / row["output_usd_per_1m_tokens"] * 1_000_000
blended_rate = (
    row["input_usd_per_1m_tokens"] * 0.75
    + row["output_usd_per_1m_tokens"] * 0.25
)
mixed_tokens_per_hour = node_hour / blended_rate * 1_000_000
print(output_tokens_per_hour, mixed_tokens_per_hour)
```

## License and commercial-use limits

Hy4 preview uses the standard Apache License 2.0. Commercial use, modification and distribution are permitted; we found no model-specific MaaS or revenue restriction in the published license. Distribution still carries duties: include the license, preserve applicable notices, mark modified files and respect the license's patent-termination terms. The license does not grant trademark rights.

That is materially simpler for a public API than a custom model license, but it is not legal advice. Your own data, fine-tunes, product claims, export controls and downstream-use obligations remain separate review items.

## Where Hy4 preview fits

- **Long-horizon software engineering:** Tencent positions it for planning, debugging, verification and front-end interaction work. Test task completion and tool-loop reliability on your repositories; vendor benchmarks are not a deployment SLO.
- **Office analysis and artifact creation:** the release targets multi-file documents, spreadsheets, presentations, equations and financial models. The production question is whether your file ingestion and tool layer preserve the relevant context.
- **Game development:** Tencent reports stronger playable-prototype and game-engine workflows across multiple turns.
- **Scientific agents:** the vendor highlights AI research, molecular dynamics, condensed-matter physics and pure mathematics. Expert verification remains necessary.
- **Serving-system research:** Gated DSA, IndexCache, iHC and the native MTP layer make Hy4 relevant to sparse-attention and speculative-decoding work.

Do not choose it for image or video input, a single-GPU deployment, or a cheap always-idle internal demo. Tencent also discloses that this early preview may reason longer than needed and over-verify its own work.

## Sources and limitations

- [Tencent Hy4 preview repository](https://github.com/Tencent-Hunyuan/Hy4-preview?utm_source=chinaapi&utm_medium=insight&utm_campaign=hy4-preview-local-deployment) and [official Tencent release](https://www.tencent.com/tencent-releases-and-open-sources-tencent-hy4-preview/?utm_source=chinaapi&utm_medium=insight&utm_campaign=hy4-preview-local-deployment) for architecture, positioning, known limitations and weight links.
- [Official BF16 checkpoint](https://huggingface.co/tencent/Hy4-preview?utm_source=chinaapi&utm_medium=insight&utm_campaign=hy4-preview-local-deployment) and [official MXFP8 checkpoint](https://huggingface.co/tencent/Hy4-preview-FP8?utm_source=chinaapi&utm_medium=insight&utm_campaign=hy4-preview-local-deployment) for repository metadata and file-size reproduction.
- [SGLang Hy4 cookbook](https://docs.sglang.io/cookbook/autoregressive/Tencent/Hy4-Preview?utm_source=chinaapi&utm_medium=insight&utm_campaign=hy4-preview-local-deployment) and [vLLM Hy4 recipe](https://recipes.vllm.ai/tencent/Hy4-preview?utm_source=chinaapi&utm_medium=insight&utm_campaign=hy4-preview-local-deployment) for hardware shapes, context sizing, kernels and serving commands.
- [Apache-2.0 license](https://github.com/Tencent-Hunyuan/Hy4-preview/blob/main/LICENSE?utm_source=chinaapi&utm_medium=insight&utm_campaign=hy4-preview-local-deployment) for commercial-use conditions.
- [Lambda on-demand instances](https://lambda.ai/instances?utm_source=chinaapi&utm_medium=insight&utm_campaign=hy4-preview-local-deployment) for the dated 8x B200 rental baseline.

We did not reproduce model quality, throughput, long-context behavior, power draw or the managed route's byte identity. The 200-person and redundant-public-service shapes are ChinaAPI capacity-planning recommendations layered on official framework configurations. Prices and preview support can change; the generated live pricing page is authoritative for ChinaAPI rates.

---
<!-- GENERATED BY scripts/gen_insights.py; locale: en; Markdown twin -->
Markdown twin of https://chinaapi.ai/insights/hy4-preview-local-deployment-value/. Full site reference: https://chinaapi.ai/llms-full.txt