Hy3 preview: Self-Hosting & Deployment Guide
Hy3 preview is Tencent's open-weight, large Mixture-of-Experts language model focused on high-efficiency reasoning and agentic workflows.
What is Hy3 preview?
Hy3 preview is a 295B-parameter Mixture-of-Experts large language model from Tencent’s Hunyuan (Hy3) family, with about 21B active parameters and an extended context window around 256K tokens. It is mainly used for complex reasoning, instruction following, long-context tasks like document analysis, and general-purpose chat or writing. It is also applied in coding, agents, and other production scenarios that benefit from configurable reasoning depth and efficient inference. Hy3 preview belongs to Tencent’s Hunyuan model line (sometimes referred to as Hunyuan 3.0), succeeding earlier Hunyuan generations such as Hy2.
How to run Hy3 preview yourself
Open weights (other), 298.8B parameters. Below: how much memory the weights need at each quantization, which hardware fits them, and the commands to serve it.
Hy3 preview VRAM requirements
| Precision / quantization | Weights | Figure | Source |
|---|---|---|---|
| BF16 (full precision) | 597.6 GB | Estimate | 298.8B params × 2 bytes |
| MLX 4-bit | 166 GB | File size | mlx-community/Hy3-preview-4bit |
File size = actual download size on Hugging Face. Estimate = parameter count × bytes per weight. Weights only — the KV cache for long contexts needs extra memory; the fit tables below add 15% headroom as a rule of thumb.
GPU requirements: deploy Hy3 preview on your hardware
| Hardware | BF16 (full precision) | MLX 4-bit |
|---|---|---|
| NVIDIA H200141 GB | 5× | 2× |
| NVIDIA H10080 GB | — | 3× |
| NVIDIA A100 80GB80 GB | — | 3× |
| NVIDIA L40S48 GB | — | 4× |
GPUs needed per variant.
| Hardware | MLX 4-bit |
|---|---|
| Mac, 512 GBM3 Ultra | Fits |
| Mac, 192 GBM2 Ultra | — |
| Mac, 128 GBM4 Max / M3 Max | — |
| Mac, 64 GBM4 Pro / Max | — |
Assumes ~75% of unified memory is usable by the GPU.
| Hardware | BF16 (full precision) | MLX 4-bit |
|---|---|---|
| AWS p5.48xlarge8× H100 · 640 GB | — | Fits |
| AWS p4de.24xlarge8× A100 80GB · 640 GB | — | Fits |
| AWS g6e.12xlarge4× L40S · 192 GB | — | Fits |
| Google Cloud a3-highgpu-8g8× H100 · 640 GB | — | Fits |
| Azure ND H100 v58× H100 · 640 GB | — | Fits |
Whole-instance GPU memory; public instance specs.
Serve it yourself
pip install -U vllm
vllm serve tencent/Hy3-previewpip install 'sglang[all]'
python -m sglang.launch_server --model-path tencent/Hy3-previewpip install -U mlx-lm
mlx_lm.generate --model mlx-community/Hy3-preview-4bit --prompt "Hello"Commands use the official repositories above; vLLM and SGLang expose an OpenAI-compatible endpoint.
Not on LLM.API yet
We do not route this model through our API at the moment. Browse the models you can call today — most workloads have a close match already live.
Providers
Different companies host the same model. OpenRouter routes your request to one of them based on the routing mode you pick — Balanced (price + speed), Nitro (fastest), or Exacto (highest tool-calling accuracy).
List price by provider
InputOutputProvider list prices; the LLM.API discount applies on top.
| Provider | Input | Output | Cache read /M | Latency | Throughput | Uptime |
|---|---|---|---|---|---|---|
| in $0.0002; out $0.0006 per 1M tokens | 262K tokens | tools, streaming, reasoning |
List prices from third-party hosts in the OpenRouter public catalogue (tencent/hy3-preview); LLM API does not serve this model and applies no discount. Updated weekly. Last updated 11 Oct 2026.
Try this model
Test Hy3 preview right here — free to start.
Suggestions for your first prompt
Not available on LLM API yet
We do not route this model through our API at the moment, so there is no endpoint or code snippet for it yet. Browse the models you can call today — most workloads have a close match already live.
Browse available modelsHy3 preview: Self-Hosting & Deployment Guide in production
What the model is good at.
-
Complex Reasoning
Performs advanced logical and mathematical reasoning, excelling on STEM tasks and challenging benchmarks and real-world exams and evaluations.
-
Instruction Following
Understands and executes nuanced natural-language instructions, with strong context learning for long prompts up to 256K tokens.
-
Agentic Workflows
Powers multi-step AI agents, integrating with frameworks like OpenClaw to orchestrate tools, search, and multi-stage task automation.
-
Code Generation
Generates and edits code, supporting complex software development workflows and scoring competitively on mainstream coding agent benchmarks.
-
Multilingual Support
Handles multiple languages, enabling cross-lingual text understanding and generation for global users across Tencent’s ecosystem and tools.
What you get when you call Hy3 preview: Self-Hosting & Deployment Guide through LLM.API.
-
Unified AI Routing
Dynamically route each request to the optimal model across providers based on cost, latency, and quality—without changing your integration or redeploying code.
One endpoint, every model. -
Cost-Aware Orchestration
Automatically balance price and performance with model-tier rules, budgets, and usage limits so you can ship rich AI features without runaway spend.
Control spend by design. -
Resilient Fallbacks
Define cascading provider and model fallbacks so requests survive outages, rate limits, and model failures with graceful degradation instead of hard errors.
No single point of failure. -
End-to-End Observability
Trace every request across models and providers with logs, metrics, and structured payloads so you can debug prompts, optimize latency, and track regressions.
See every token, everywhere. -
Task-Level Abstractions
Describe tasks like chat, generation, tools, or scoring once and let LLM.API map them to provider-specific APIs so you avoid vendor lock-in glue code.
Code to tasks, not vendors. -
High-Throughput Batch
Submit massive batches of prompts, evaluations, or embeddings through a single pipeline optimized for concurrency limits, retries, and cost-efficient parallelization.
Scale experiments, not ops.
When to Use — When NOT to Use
Use it if...
- You need a Chinese-developed model suitable for deployment within Tencent’s cloud ecosystem.
- You need a general-purpose assistant for chat, drafting, and everyday productivity tasks.
- Your use case involves experimenting with a newer Tencent model in preview environments.
- You need to prototype multilingual chatbots primarily targeting Chinese and English users.
- Your use case involves integrating with other Tencent services or existing Tencent infrastructure.
Avoid if...
- You need a fully production-hardened model with long-term stability beyond a preview phase.
- Your workload requires guaranteed enterprise SLAs, compliance attestations, and audited certifications.
- You need state-of-the-art long-context reasoning over very large documents or codebases.
- Your workload requires rich ecosystem tooling, plugins, and community resources already battle-tested.
- You need proven performance benchmarks against leading frontier models for mission-critical decisions.
COMMUNITY
What developers say about Hy3 preview
Summarised from public developer discussions and write-ups. Opinions are the sources’, not LLM.API’s.
- Developers on Hacker News noted the model's ability to generate interactive HTML features like functional buttons, though some reported inconsistencies where requested animations failed or targeted the wrong UI elements.
- A developer at Engr Mejba Ahmed observed that Hy3 Preview outperforms larger open-weight models in most areas except coding, where it remains competitive but slightly behind the current leaders.
- Developers on Hacker News reported that while the model successfully follows complex instructions, it may occasionally fail to render specific graphics formats, such as confusing HTML with SVG.
SOURCES
Frequently Asked Questions
-
What is Hy3 preview?
Hy3 preview is a Tencent large language model accessible via LLM.API, suitable for general-purpose text generation and analysis tasks.
-
What is Hy3 preview best suited for?
Hy3 preview is best for fast, low-friction chat-style completion, coding assistance, and structured text generation where low latency matters.
-
What is the context window of Hy3 preview?
Hy3 preview supports a mid-sized context window, suitable for multi-turn conversations, medium-length documents, and typical code files.
-
What modalities does Hy3 preview support?
Hy3 preview currently supports text input and text output only when accessed through LLM.API.
-
How is Hy3 preview priced on LLM.API?
Hy3 preview uses LLM.API’s unified per-token pricing; you are billed for input and output tokens at the Tencent Hy3 preview rate.
-
How fast is Hy3 preview in terms of latency and throughput?
Hy3 preview typically returns initial tokens quickly and supports streaming, making it suitable for interactive applications and tooling.
-
How do I call Hy3 preview through the LLM.API gateway?
Specify the model name "tencent-hy3-preview" (exact name may vary) in your LLM.API request and provide your LLM.API key for authentication.
-
How does Hy3 preview compare to other similar models on LLM.API?
Compared to larger flagship models, Hy3 preview generally trades some reasoning depth and creativity for lower cost and faster responses.
-
What are the main limitations of Hy3 preview?
Hy3 preview can hallucinate facts, struggle with very long contexts or complex reasoning chains, and should not be used without human review for high-stakes decisions.
-
Can I fine-tune or customize Hy3 preview via LLM.API?
Direct fine-tuning is typically not available; instead, you customize Hy3 preview behavior using system prompts, few-shot examples, and application-side orchestration.
COMPARE
Competitive Models
-
Claude Sonnet 4.5
Up to 30%
Claude Sonnet 4.5 is an Anthropic large language model optimized for software development, computer use, and agentic workflows, offering strong performance on coding and reasoning tasks at mid-tier pricing. It is part of the Claude 4.5 generation and is available through multiple cloud providers and enterprise platforms.
-
GLM 5.1
GLM 5.1 is Z.ai’s flagship open-weight Mixture-of-Experts large language model optimized for long-horizon agentic coding and software engineering tasks. It is notable for its very large context window, strong SWE-Bench Pro performance, and open-source MIT licensing.
-
GPT-5.2
Up to 30%
GPT-5.2 is an OpenAI large language model in the GPT-5 family, designed for advanced natural language understanding and generation across many tasks. It emphasizes improved reasoning, safety, and versatility compared with earlier GPT models.
Get one key to every model
Swap your API key. Keep your code.