Tencent Hy3 is available on Novita AI as a serverless API at no cost — $0 per million tokens for input and output as of July 2026. It is a 295B Mixture-of-Experts model with 21B active parameters per token, a 256K token context window, and three configurable reasoning modes. Developers who need a capable open-weight model for coding agents, reasoning workflows, or long-context tasks now have a straightforward path: use model ID tencent/hy3 against the Novita AI endpoint and start without committing to paid infrastructure.
Key Takeaways
- Free at launch: Hy3 is currently priced at $0/M tokens for both input and output on Novita AI.
- 295B total, 21B active: MoE architecture means only 21B parameters fire per token, keeping inference costs low even as capability scales.
- 256K context window and 262K max output: Large enough for full repository snapshots, long agent transcripts, and sustained reasoning chains.
- Three reasoning modes: Switch between
no_think(fast),low, andhigh(deep chain-of-thought) using thereasoning_effortparameter. - Strong coding and agent benchmarks: 74.4% on SWE-bench Verified and 87.2% on GPQA Diamond — competitive with the best open-source models available today.
What Is Hy3?
Hy3 is Tencent’s third-generation large language model, developed by the Tencent Hy team (formerly branded as Hunyuan). The model was first released as Hy3-preview on April 23, 2026, followed by the full Hy3 release in July 2026. Both are open-sourced under an Apache 2.0 license with no geographic restrictions — an improvement over the preview’s community license that originally excluded the EU, UK, and South Korea.
The model is built on a rebuilt pre-training and reinforcement learning infrastructure that Tencent rearchitected from scratch in early 2026. That rebuild is the main reason the jump from Hy2 to Hy3 is as large as it is: SWE-bench Verified went from 53.0% to 74.4% — roughly a 40% relative improvement.
Hy3 powers Tencent’s Yuanbao assistant and CodeBuddy developer product. It was designed with multi-step agentic workflows as a first-class target, not an afterthought.
Architecture in brief
The model uses a dense-MoE hybrid design. The first transformer layer is dense (all parameters active); layers 2 through 80 are sparse MoE. Each MoE layer contains 192 routed experts and 1 always-active shared expert, with each token activating a top-8 subset via sigmoid gating. The result is a model that runs like a 21B-parameter model at inference time while drawing on 295B parameters of stored knowledge.
A separate 3.8B Multi-Token Prediction (MTP) layer enables speculative decoding, which improves throughput in serving environments.
Attention uses Grouped Query Attention (GQA) with 64 query heads and 8 key-value heads, per-head QK RMSNorm, and RoPE positional encoding.
Hy3 API Access on Novita AI
Hy3 is available on Novita AI as a serverless chat model with an OpenAI-compatible API. No GPU provisioning, no instance management, no minimum spend — you call the endpoint and pay only for what you use (currently $0 per million tokens).
The Novita AI model library has a playground for testing before building an integration.
The endpoint is OpenAI-compatible: set base_url to https://api.novita.ai/openai, set model to tencent/hy3, and use your Novita API key. The endpoint supports function calling, structured outputs, and reasoning controls. For existing OpenAI SDK integrations, the only change is swapping base_url and model.
Hy3 Specs and Pricing on Novita AI
| Field | Details |
|---|---|
| Display name | Hy3 |
| Model ID | tencent/hy3 |
| Base URL | https://api.novita.ai/openai |
| Endpoint family | chat/completions |
| Context window | 262,144 tokens |
| Max output tokens | 262,144 tokens |
| Input pricing | $0 / M tokens |
| Output pricing | $0 / M tokens |
| Supported features | Function calling, structured outputs, reasoning |
| Architecture | 295B total / 21B active, MoE |
| Best fit | Coding agents, reasoning, long-context analysis |
The $0/M pricing is a launch-period rate. The Novita AI pricing page has current rates.
Hy3 Benchmark and Performance Signals
Tencent published benchmark results for the Hy3-preview release; the full Hy3 model represents an additional post-training improvement.
| Benchmark | Hy3-preview Score | Context |
|---|---|---|
| SWE-bench Verified | 74.4% | vs. 53.0% for Hy2; top-tier open-weight result as of April 2026 (Tencent, Artificial Analysis) |
| GPQA Diamond | 87.2% | Graduate-level scientific reasoning (Tencent GitHub) |
| MMLU | 87.42% | Broad academic knowledge (Tencent GitHub) |
| Terminal-Bench 2.0 | 54.4% | Real-world software engineering tasks (Tencent GitHub) |
In a blind evaluation with 270 domain experts completing real work tasks, Hy3-preview scored 2.67/4 versus 2.51/4 for GLM5.1. Strongest result categories were front-end development, data and storage, and CI/CD.
The practical ceiling to understand: Hy3 trails the top closed-source models in head-to-head comparisons tracked on Artificial Analysis. It does not support image input, which rules it out for multimodal tasks. Within the open-source and open-weight tier, it is one of the stronger options for coding and agentic work as of mid-2026.
Benchmarks for the full Hy3 release were not separately published at time of writing; Tencent describes the full model as improving over the preview in reasoning, agentic workflows, and long-context tasks following feedback from 50+ product teams.
Key Capabilities for Developers
Coding and software engineering. Hy3 was trained with agentic code workflows as a primary target. The 74.4% SWE-bench score reflects that focus — the model can reason about multi-file changes, identify bugs in context, and produce coherent diffs across large codebases. It performs especially well in front-end development and CI/CD scenarios based on the expert evaluation data.
Long-context reasoning. A 256K context window means you can pass in full repository snapshots, complete API documentation, long agent transcripts, or multi-document analysis without trimming. The 262K max output gives room for long plans, detailed explanations, and generated documentation.
Configurable reasoning depth. Three reasoning modes let you tune latency versus quality per request. Use no_think for fast routing decisions or simple completions; use high for hard agentic steps that need deep chain-of-thought. This is useful when building agent pipelines where not every call warrants full reasoning compute.
Function calling and structured outputs. Both are supported on Novita AI, which means Hy3 fits into standard tool-use patterns without adapter shims.
Multilingual. The model handles Chinese and English well, which matters for teams working with Chinese-language data, users, or documentation alongside English systems.
When to Use Hy3
Coding agents and automated software workflows. SWE-bench 74.4% places Hy3 close to the current open-source ceiling for software engineering tasks. If your workflow involves identifying and fixing bugs, navigating large codebases, or generating code from specifications, Hy3 is a strong candidate to test.
Long-context analysis where cost matters. The $0/M token launch pricing, combined with 256K context and MoE efficiency, makes it unusually affordable for long-document tasks. If you are chunking documents to fit into smaller models because of cost constraints, Hy3 lets you test end-to-end long-context reasoning at zero inference cost during this period.
Multi-step agentic pipelines. The model’s design was shaped by real-world agent deployments across Tencent’s products. The configurable reasoning modes are practical for orchestrating pipelines where you want fast responses on simple steps and deep reasoning on complex ones.
Open-weight requirements. Apache 2.0 licensing means the model weights are available for inspection, fine-tuning, and self-hosting. If your team needs to audit the model or run it in a controlled environment, the hosted API and the self-hosted path coexist.
When Not to Use Hy3
Multimodal tasks. Hy3 is text-in, text-out. It does not accept image, audio, or video inputs. For vision tasks, look at models like GLM-4.6V, ERNIE 4.5 VL, or the Hunyuan image model family.
Tasks requiring sub-second latency at scale. MoE routing adds overhead compared to dense models of equivalent active parameter count. For high-volume, latency-sensitive production use cases, benchmark actual response times under realistic load before committing.
When you need the absolute frontier. Hy3 is competitive for an open-source model but trails the top closed-source models on independent benchmarks tracked by Artificial Analysis. If your task critically depends on the highest possible reasoning quality, evaluate those differences in your specific workload.
Embedding or classification tasks. Hy3 is a generative chat model. It is not an embedding or reranking model. Use a dedicated embedding model for semantic search and retrieval.
How Hy3 Fits Your API Workflow
The Novita AI endpoint uses the standard OpenAI chat completions schema. If you already call any OpenAI-compatible model through Novita, the integration change is a model ID swap from your current model to tencent/hy3.
The reasoning_effort parameter controls which thinking mode fires on a given request. Setting it to "high" enables deep chain-of-thought and is appropriate for hard agent steps. Setting it to "none" or "low" gives faster responses at the cost of reasoning depth. This parameter is not standard OpenAI spec — the Novita AI docs have the exact field name and accepted values.
Function calling follows standard OpenAI tool-use format. Structured outputs work via response_format. Both let you use Hy3 as a drop-in replacement inside existing agent frameworks (LangChain, LlamaIndex, CrewAI, etc.) without changing tool definitions or output parsers.
For rate limits and context budget management at production scale, the Novita AI LLM API documentation covers throughput details and request patterns.
Conclusion
Hy3 is the right model to benchmark if you are running coding agents, long-context analysis workflows, or agentic reasoning pipelines and have not yet found an open-weight model that matches your quality bar. The combination of 74.4% SWE-bench, 256K context, configurable reasoning depth, and $0/M launch pricing on Novita AI removes most of the friction from trying it: add the model ID, run your existing eval, and measure the delta before deciding whether to route production traffic.
The ceiling is real — it does not beat the best closed models, and it does not handle vision. But for the use cases it targets, it is among the stronger open-source options available today, and the Novita AI serverless path means you can test it without a GPU provisioning commitment. The Hy3 model page on Novita AI includes a playground for testing before building an integration.
FAQ
What is the model ID for Tencent Hy3 on Novita AI?
The exact model ID is tencent/hy3. Use this in the model field of your API calls.
Is Hy3 free on Novita AI?
Hy3 is currently priced at $0 per million tokens for both input and output on Novita AI. The pricing page has current rates, as launch pricing may change.
What is the context window for Hy3?
Hy3 supports a 262,144-token context window with a 262,144-token max output on Novita AI.
Does Hy3 support function calling?
Yes. The Novita AI listing for Hy3 includes function calling, structured outputs, and reasoning as supported features.
What reasoning modes does Hy3 support?
Hy3 includes three reasoning modes: no_think (fast, no chain-of-thought), low (light reasoning), and high (deep chain-of-thought for complex tasks). These are configured via the reasoning_effort parameter.
Does Hy3 support image input?
No. Hy3 is a text-only model. It does not accept image, audio, or video inputs.
How does Hy3 compare to Kimi K2 or GLM 5.2?
All three are strong open-weight MoE models aimed at coding and agentic workflows. Kimi K2 has 1T total parameters with 32B active and leads in math and STEM benchmarks. GLM 5.2 has a 1M-token context window optimized for long-context repository analysis. Hy3 sits at 295B/21B with strong SWE-bench results (74.4%) and a free launch period on Novita AI. The right choice depends on your specific task and context requirements.
Is Hy3 the same as Hunyuan?
Hy3 is developed by the same Tencent team that built the Hunyuan model series. The team rebranded from “Hunyuan” to “Hy” with this generation. Hy3 is not related to HunyuanVideo or Hunyuan Image, which are separate media generation models.
