Hy3 (Tencent Hunyuan 3) is the third generation hybrid expert (MoE) large language model launched by Tencent Hunyuan team. It has a total parameter of 295B, an activation parameter of 21B, 192 expert Top-8 routes, 256K context windows, and an open source Apache 2.0 license. It reaches 74.4% on SWE-bench Verified and 90.4% on GPQA Diamond. It has powerful code generation, tool calling and agent capabilities, and has been applied to 50+ Tencent internal products.

Hy3 (Tencent Hunyuan 3): The benchmark of Agent and code capabilities of the open source MoE large model

Core parameters and statistics of Hy3

Hy3 (Hunyuan 3) is the third-generation Mixture-of-Experts (MoE) large language model launched by Tencent’s Hunyuan team. It is also Tencent’s strongest and most open basic model to date. Compared with the previous generation Hy2, Hy3 has achieved cross-generation upgrades in parameter magnitude, expert scale, context length and reasoning capabilities. It is fully open source under the Apache 2.0 license and has no geographical restrictions.

Project	Specifications
Product name	Hy3 (Hunyuan 3 / Tencent Hunyuan third generation)
Product Type	Basic Large Model (MoE)
Total parameters	295B (299B including MTP layer)
Activation parameters	21B
MTP layer parameters	3.8B
Number of experts	192 experts, Top-8 routing
Number of network layers	80 layers (excluding MTP)
Attention heads	64 (GQA, 8 KV heads, head dim 128)
Hidden layer dimension	4096
Middle layer dimension	13312
Context window	256K tokens
Vocabulary size	120,832
Support precision	BF16/FP8
License Agreement	Apache 2.0
Delivery form	Open source weight (HuggingFace / ModelScope) + API (OpenRouter and other third parties)
Home	CN (Tencent)
Supported languages	Chinese, English
Release Date	Preview: 2026-04-23; Full: 2026-07-06

Hy3 is not positioned as an omniscient and omnipotent general model, but rather "pursues the highest inference quality under a specific parameter budget". The 21B activation parameters only require about 4.2% of the network to participate in the calculation (8/192 experts) when running inference, which makes its deployment cost much lower than dense models of the same intelligence level. The 256K context window can process approximately 384 pages of A4 documents at one time, which is suitable for long document analysis and cross-file code base understanding.

Actual meaning of efficiency improvement: Compared with other models with 21B activation parameters and 100B+ activation parameters, Hy3 has a natural advantage in inference throughput. When deployed with vLLM or SGLang, the memory usage and token generation delay of a single inference are lower than dense models of the same intelligence level. For high-frequency API callers, this is directly reflected in lower unit token prices and lower batch processing latency.

Reasoning mode design: Supports reasoning_effort three levels of control - no_think (straight-out mode, suitable for deterministic tasks such as translation and summarization), low (light thinking, suitable for regular question and answer), high (deep chain reasoning, suitable for mathematical proofs, competition programming, complex Agent planning). When high mode is selected, the amount of token output will increase by 3-8 times, and the cost of a single call will increase accordingly.

Users and market recognition of Hy3

Hy3's market recognition shows a two-wheel drive pattern of "Tencent's internal production verification + open source community reputation".

Internal production scale: Hy3 has been integrated into Tencent’s 50+ internal product lines before its official release, covering core businesses such as WeChat, QQ, Tencent Cloud, and game platforms. This level of internal deployment means that the model has been tested by real traffic and complex business logic - including multi-round dialogue stability, long context attenuation control, multi-language mixing and other engineering challenges that have been polished in the production environment.

Open source community response: As of July 2026, the Tencent-Hunyuan/Hy3 warehouse on GitHub has received 530+ Stars, 120+ Forks, and the monthly downloads of model weights on HuggingFace exceed 14,000 times. The community around Hy3 has produced 10+ fine-tuned variants and 65+ quantized versions (supporting tool chains such as llama.cpp, LM Studio, Jan, etc.), and 15 HuggingFace Spaces provide online experience entrances.

Third-party evaluation performance: On Artificial Analysis's Intelligence Index, Hy3 ranked 8th among ultra-large parameter scale open source models with a score of 41 (a total of 97 model comparisons), significantly higher than the category median of 25 points. Key benchmark scores include:

• GPQA Diamond: 90.4% (knowledge reasoning, top level of open source model)
• SWE-bench Verified: 78% (actual coding task solution rate)
• SWE-bench Pro: 57.9% (complex engineering tasks)
• Humanity's Last Exam: 53.2% (comprehensive knowledge challenge)
• Deep-SWE: 28% (independent code repair)
• Apex Agents: 25.6% (Agent tasks)
• LHTB (long-term terminal task): 28.8% (long-term Agent task)

Blind test expert evaluation: Tencent jointly conducted a blind test with 270 industry experts. Hy3 received an overall score of 2.67/4, exceeding GLM-5.1's 2.51/4. The biggest advantages are concentrated in front-end development, data storage and CI/CD tasks. This result better reflects the model utility in real scenarios than a single benchmark.

Cost Advantages of Hy3

Hy3's cost structure consists of three layers: "Apache 2.0 open source zero license fee + low activation parameter reasoning efficiency + low price of third-party API". It is one of the most cost-effective open source large models currently.

C client/individual users: Directly download model weights through HuggingFace (tencent/Hy3), completely free. Individual developers can run inference on their own hardware without paying any licensing or API fees. However, you need to bear the cost of GPU computing power yourself - to run Hy3 full-precision inference with 8×H20-80G or equivalent configuration, a single hardware investment is in the order of tens to hundreds of thousands of yuan. For occasional usage scenarios, it is more economical to pay as you go for third-party APIs.

Developer/API call (taking OpenRouter as an example):

Billing items	Price (USD/million tokens)	Comparison reference
Input (cache hit)	$0.035 (-75% discount)	Claude Sonnet 5: $3/M
Input (standard)	$0.14	GPT-5.6: ~$15/M
Output	$0.58	Grok 4.5: $6/M
Mixed price (7:2:1 ratio)	Approx. $0.11	DeepSeek V4 Flash: Approx. $0.08

Hy3’s API price is approximately 1/10 to 1/50 of that of closed-source models of similar intelligence levels. For a lightweight application that processes 1 million token inputs and 500,000 token outputs per day, the monthly cost is approximately ($0.14 × 30 + $0.58 × 15) = $12.9/month, which is almost negligible.

Enterprise/Private Deployment: The open source weight is free (Apache 2.0), but the infrastructure cost - taking an 8×H20-80G server as a reference, the monthly rental cost of a single unit is about $8,000-12,000 (Evian manufacturer and contract length fluctuate), which can support moderately concurrent Hy3 inference. The total cost of the enterprise should be evaluated based on the "API annual fee vs private three-year cost TCO" model, and the hidden costs of operation and maintenance manpower, model version iterative updates and quantitative deployment (FP8 further reduces video memory requirements) need to be included.

Purchasing Tips: API pricing is based on the real-time page of each service provider. There are price differences in different regions (such as mainland China vs overseas). The recommended hardware configuration and performance indicators for private deployment are based on the Deployment chapter of the GitHub repository.

Main functions of Hy3

The functional design of Hy3 is centered around the four core capability lines of "Code Generation + Agent Tool Calling + Deep Reasoning + Long Context".

• Code generation and engineering tasks: Covering code generation, completion, refactoring, bug diagnosis and unit test generation for mainstream languages ​​such as Python, JavaScript/TypeScript, Java, and Go. A 78% solution rate on SWE-bench Verified means the model is able to understand real GitHub Issue descriptions, locate relevant code files, and propose mergeable fixes. Implementation Tips: For complex engineering tasks, it is recommended to enable the reasoning_effort=high mode. In the straight-out mode, boundary condition processing may be missed.

• Agent Tool Calling (Function Calling): Hy3’s Tool Calling capability is the core highlight of this upgrade. The model natively supports multiple rounds of tool calls, error recovery, and cross-Agent framework generalization—SWE-bench Verified accuracy fluctuates within 4% under different scaffolds (CodeBuddy, Cline, KiloCode). Tool open list: Supports standard tool behaviors such as navigate (web page navigation), click (click element), type/input (text input), screenshot (page screenshot), extract (information extraction), wait (waiting conditions), read_file/write_file (file reading and writing), execute_command (command execution), finish (task end report). Architecture link: LLM (Hy3) → API/MCP Server → Tool Execution Environment → Observation returns LLM, forming an interactive closed loop of "Planning → Execution → Observation → Adjustment".

• Deep Reasoning and Thought Chain: Supports adjustable chain-of-thought reasoning. Reaching the top level of open source models in knowledge question answering (GPQA Diamond 90.4%) and complex reasoning tasks (Humanity's Last Exam 53.2%). Internal reviews show that Hy3’s hallucination rate dropped from 12.5% ​​in Hy3 Preview to 5.4%, and its common sense error rate dropped from 25.4% to 12.7%, a significant improvement in factual accuracy and logical consistency.

• Long context understanding (256K tokens): Can process about 384 pages of text content at a time, suitable for scenarios such as legal contract set review, global analysis of large code bases, long conversation history understanding, etc. Through the joint optimization of SFT and RL, the rate of referential resolution, omission recovery, and constraint inheritance problems in multi-round conversations is reduced from 17.4% to 7.9%. Implementation Tips: Inference latency and KV cache overhead will increase significantly in long context scenarios. It is recommended to enable full context in tasks that really require cross-chapter correlation, and control daily tasks within 32K to balance costs.

• Multi-modal extension: The core of Hy3 is a text model, but it can be extended to image and text understanding through Tencent's visual language model pipeline. The native Hy3 weighted visual mode is not supported, and plain text scenes do not need to pay attention to this limitation.

• Engineering optimization tool chain: Provides a complete fine-tuning (Finetune) pipeline (based on vLLM/SGLang), RL post-training (GRPO via verl + Megatron-LM + vLLM rollout) and quantization tools (AngelSlim supports common quantization algorithms, low-bit quantization, and speculative sampling). Developers can build vertical domain-specific models based on Hy3 weights.

Hy3 model and version evolution

The version iteration of Hy3 follows the rhythm of "Preview verification → Feedback collection → Full Release optimization".

Mainline release

• Hy2 (~2025-12): Hunyuan second-generation model, SWE-bench is about 53.0%, laying a baseline for Hy3’s MoE architecture upgrade. There is no official precise release date yet.
• Hy3 Preview (2026-04-23): The first batch of training models on the new infrastructure, using the 295B MoE (21B activation) architecture and 192 expert Top-8 routing for the first time. SWE-bench Verified reached 74.4%, attracting the attention of the developer community. After the release, Tencent collected 50+ internal product usage feedback.
• Hy3 Full (2026-07-06): Large-scale post-training optimization based on Preview feedback. Key improvements include:
  • Tool calling stability and error recovery capabilities have been significantly improved
  • Illusion rate reduced from 12.5% to 5.4%, common sense error rate reduced from 25.4% to 12.7%
  • Multi-turn conversation intent tracking issue rate reduced from 17.4% to 7.9%
  • The output is more concise while keeping complex intentions from decaying in long-distance interactions
  • Simultaneous release of FP8 quantitative version (Hy3-FP8) to lower the deployment threshold

Candidate Verification

• Hy3-FP8 (released on the same day as Full): A quantized version of FP8 that works with the AngelSlim tool chain, which greatly reduces the graphics memory requirements while maintaining the quality of inference, allowing more developers to run Hy3 on limited hardware.

Release Notes: Hy3's open source policy is released under the Apache 2.0 license, with no additional terms and no geographical restrictions. The weights of the old version (Hy2) have also been made public via the Tencent HuggingFace organization page. Subsequent version iteration plans are subject to information released by the GitHub warehouse and official Twitter/X account.

Technical advantages of Hy3

Hy3’s technical route selection reflects Tencent’s engineering accumulation in “maximizing reasoning quality under limited computing power budget”.

MoE architecture and 192 expert Top-8 routing: Hy3's MoE does not simply increase the number of experts, but deploys 192 expert subnetworks per layer in the 80-layer Transformer, and only 8 of them are activated per token (about 4.2% activation rate). This means that only 21B of the 295B parameters participate in each step of reasoning, storing extensive knowledge while controlling the amount of reasoning calculations. Different tokens will be routed to different experts - code tokens tend to activate code expertise experts, and natural language tokens activate language expertise experts, forming an implicit task partition.

MTP layer (Multi-task prediction layer): The shared MTP layer of 3.8B parameters coordinates the output of different experts, ensuring that even if different tokens are processed by different experts, the overall sequence maintains semantic coherence. This is the key mechanism for Hy3 to maintain stable output in complex multi-step reasoning.

Reasoning Effort: The model can be dynamically adjusted between "speed" and "depth" through the reasoning_effort parameter. In no_think mode, the model produces answers directly, which is suitable for delay-sensitive scenarios; in high mode, the model develops a complete chain reasoning. This design allows the same model weights to be adaptively adjusted under different delay budgets without the need to deploy multiple versions.

Post-training pipeline (SFT + RL): Hy3’s SFT and GRPO reinforcement learning pipeline runs on the verl + Megatron-LM framework, supporting large-scale distributed training. The improvement in post-training data quality and diversity is the main reason for the jump in quality from Hy3 Preview to Full version. Internal evaluation shows that the joint optimization effect on hallucination rate, common sense errors and multi-round consistency is significant.

Production-grade tool call stability: Hy3 fixes multiple baseline reliability issues to bring tool calls up to production-grade standards under different configurations and output constraints. The variance of the SWE-bench Verified accuracy of the model on different Agent frameworks (CodeBuddy, Cline, KiloCode) remains within 4%, indicating that the tool calling capability is not bound to a specific framework implementation.

Engineering Deployment Adaptation: Hy3 has official adaptation solutions on the two mainstream inference engines vLLM and SGLang, supporting MTP speculative decoding acceleration. It is recommended to use 8×H20-80G or higher memory GPU during deployment. The FP8 quantized version can further reduce memory requirements. The quantification tool AngelSlim is open sourced by Tencent and supports common quantization algorithms and low-bit quantization.

Performance and Throughput Reference: According to actual measurement by Artificial Analysis, the median output speed of Hy3 on the third-party API is about 60.3 token/s (lower than the category median of 64.2, but the gap is not big), and the median TTFT (first word delay) is about 2.65 seconds (slightly higher than the category median of 1.83 seconds). The specific values ​​vary depending on the deployment hardware, quantification method and concurrency, and are subject to actual testing.

Hy3’s adaptation boundaries and scene constraints

Best in: Code generation and Agent tasks, mathematical reasoning and knowledge question and answer, Chinese/English bilingual processing, tool calling and multi-step planning.

Not good at: The fineness of English document style is slightly lower than that of pure English training models; the familiarity with long-tail Western frameworks and tools is not as good as that of more ecologically mature models such as DeepSeek; the visual modality is not supported (pure text model); the reasoning delay increases significantly in ultra-long context (close to the 256K limit) scenario.

Comparison with competing products:

Comparison	Hy3	DeepSeek V4 Pro	GLM-5.2	Claude Sonnet 5
Total parameters	295B	1.6T	About 744B	Undisclosed
Activation parameters	21B	About 49B	About 100B+	Undisclosed
Context	256K	1M	~256K	~200K
License	Apache 2.0	MIT	MIT	Closed Source
API input price	$0.14/M	¥1-3/M	$0.5-1/M	$3/M
SWE-bench	78%	about 75%	about 62%	about 70%

The core advantage of Hy3 is to achieve the strongest SWE-bench performance with the smallest activation parameters and the fully open license of Apache 2.0. The disadvantage is that the ecological maturity is not as good as DeepSeek, and the 1M ultra-large context scene has not yet been covered.

How to use Hy3

Hy3 provides two usage modes: "open source self-hosting + third-party API", covering all scenarios from personal experiments to enterprise-level deployment.

How to use	Suitable for people	Features	Cost
Third-party API (OpenRouter, etc.)	Developers, small teams	Ready-to-use, no GPU required	Pay-as-you-go ($0.14/$0.58 per M token)
Self-hosting (vLLM)	Enterprise, technical team	Data does not leave the domain, full control	GPU computing cost
Self-hosted (SGLang)	Enterprise, technical team	Official adaptation, MTP acceleration	GPU computing power cost
HuggingFace weight download	Researchers, model developers	Fine-tuning/quantification/secondary development	Free (need to bring your own hardware)

API Quick Start (OpenRouter compatible with OpenAI SDK):

from openai import OpenAI

client = OpenAI(
    base_url="https://openrouter.ai/api/v1",
    api_key="<YOUR_API_KEY>",
)

response = client.chat.completions.create(
    model="tencent/hy3",
    messages=[
        {"role": "user", "content": "Use Python to implement a cached Fibonacci sequence calculation."}
    ],
    temperature=0.9,
    top_p=1.0,
    extra_body={
        "chat_template_kwargs": {"reasoning_effort": "high"}
    }
)
print(response.choices[0].message.content)

Key parameter description: temperature is recommended to be 0.9, top_p is recommended to be 1.0; reasoning_effort is optional no_think (straight out), low (light thinking), high (deep reasoning). It is recommended to use high mode for code and complex reasoning scenarios.

Self-hosted - vLLM deployment (8 GPU recommended):

# Need to compile and install vLLM first (based on source code)
export VLLM_FLASHINFER_ALLREDUCE_BACKEND=trtllm
vllm serve tencent/Hy3 \
  --tensor-parallel-size 8 \
  --speculative-config.method mtp \
  --speculative-config.num_speculative_tokens 2 \
  --tool-call-parser hy_v3 \
  --reasoning-parser hy_v3 \
  --enable-auto-tool-choice \
  --port 8000 \
  --served-model-name hy3

Self-Hosted - SGLang Deployment:

python3 -m sglang.launch_server \
  --model tencent/Hy3 \
  --tp-size 8 \
  --tool-call-parser hunyuan \
  --reasoning-parser hunyuan \
  --speculative-num-steps 2 \
  --speculative-eagle-topk 1 \
  --speculative-num-draft-tokens 3 \
  --speculative-algorithm EAGLE \
  --port 8000 \
  --served-model-name hy3

After deployment is complete, local services can be called through the OpenAI compatible interface. Detailed deployment documents are subject to the README of the GitHub repository Tencent-Hunyuan/Hy3.

Product Pricing for Hy3

Hy3’s pricing continues the “open source weight free + third-party API pay-as-you-go” model.

Open source weight: Apache 2.0 license, no usage restrictions, no geographical restrictions, no commercial use caps. Users are free to download, use, modify and redistribute it. Unlike Meta's Llama community license, Hy3 does not require additional application for commercial license.

Third-party API pricing: Based on the median price on OpenRouter, enter $0.14/million tokens and output $0.58/million tokens. 75% discount on cache hits ($0.035/million tokens). The prices of different service providers (such as Fireworks AI, Together AI, etc.) may be slightly different, and are subject to the real-time page of each platform.

Private deployment cost: The weights are free, but the infrastructure costs need to be borne by yourself. It is recommended to configure 8×H20-80G or equivalent GPU, and the monthly rental cost of a single unit is about $8,000-12,000. The FP8 quantitative version (Hy3-FP8) can carry higher concurrency on the same hardware, or lower the hardware threshold to a single card A100-80G (experimental scenarios only).

Enterprise Volume Purchase: Tencent has not disclosed official hosting API pricing for Hy3. For large traffic scenarios, it is recommended to contact the Tencent team through [email protected] to obtain business solutions, or negotiate discounts based on enterprise-level contracts on platforms such as OpenRouter.

Hy3 application scenarios

Hy3's application scenarios span four dimensions: software development, agent automation, knowledge-intensive analysis, and enterprise productivity improvement.

• Software development (code generation and engineering tasks): This is the strongest scenario for Hy3. Covers code completion, bug location and repair, unit test generation, code review, refactoring suggestions, technical document writing, etc. SWE-bench Verified performance of 78% means Hy3 is able to understand complex GitHub Issue descriptions and generate mergeable fix code. Implementation Tips: It is recommended to use it in conjunction with Agent frameworks such as CodeBuddy, Cline or KiloCode to give full play to its tool calling capabilities. Enable reasoning_effort=high mode for complex engineering tasks.

• AI Agent and Automated Workflow: Hy3's Tool Call stability has been verified by 50+ Tencent internal products. Suitable for web page automation (information collection, form filling), command line tool chain arrangement, multi-step data processing pipeline and other scenarios. Architecture Link: LLM(Hy3) → Tool Parser → Function Execution → Observation → LLM Re-planning. Guide to engineering pitfalls:

  1. Infinite loop and Token inflation control: Limit the maximum number of execution rounds through the max_steps parameter (10-20 is recommended), and set a single-step timeout, in conjunction with repeated action detection (terminate when the same Observation is returned for 3 consecutive steps) to prevent the model from idling.
  2. DOM/Context Overload: For web page Agent tasks, it is recommended to disable screenshot delivery and only return the accessibility tree or DOM summary to prevent the entire HTML document from filling the context window.
  3. Security and Ultra-Authority Governance: Set a confirmation point (Human-in-the-loop) for irreversible operations (deleting files, sending messages, paying, and publishing), or start in read-only mode in the execution environment to prevent misoperations from the model.

• Knowledge Questioning and Document Analysis: GPQA Diamond’s performance of 90.4% shows that Hy3 is close to cutting-edge closed-source models on knowledge-intensive tasks. The 256K context makes it suitable for scenarios such as reviewing long academic papers, comparing legal contract terms, and summarizing technical white papers. Implementation Tips: In document analysis scenarios, just use reasoning_effort=no_think or low mode. There is no need to enable deep reasoning, which helps reduce latency and token consumption.

• Enterprise Productivity (Internal Tools and RAG Systems): Combining open source weighted self-hosting and the RAG (Retrieval Augmented Generation) framework, enterprises can build private knowledge question answering systems. Apache 2.0 licensing ensures that data does not leave corporate infrastructure and there are no privacy risks from third-party API calls. Implementation Tips: For multi-lingual mixed document scenarios, it is recommended to cooperate with Hybrid Search (keyword + semantic retrieval) and Rerank mechanism to improve recall quality. Hy3 itself does not provide document-level RBAC, and permission isolation needs to be implemented at the RAG framework layer.

Not suitable for scenarios: Brand copywriting that requires extreme creativity (Hy3's output style is more engineering-oriented, and the creative richness is not as rich as the Claude series); non-coding scenarios that require extremely high English document style (such as academic paper polishing); scenarios that require multi-modal input (such as image recognition, video understanding).

Applicable groups of Hy3

• Software Developers & Engineers: Hy3’s core audience. Whether it’s daily coding assistance, bug diagnosis, code review or complex engineering tasks, Hy3’s 78% performance on SWE-bench means it can become a reliable programming partner for developers. Not suitable for boundaries: For teams that have extremely high quality requirements for English document generation (such as English documents for open source projects), it is recommended to use tools such as Grammarly for secondary polishing.

• AI Agent Developers and Automation Engineers: Developers who need to build automated workflows - from web page data collection to command line tool chain orchestration, Hy3's Tool Call stability and cross-framework generalization capabilities make it an ideal base model for Agent systems. Prerequisites: You need to configure the Agent framework (such as Cline, CodeBuddy) and set a reasonable step budget and confirmation point strategy.

• Enterprise & Technology Team (Private Deployment): Enterprises focused on data sovereignty and compliance. Hy3’s Apache 2.0 licensing and self-hosting capabilities make it ideal for privatized AI infrastructure. Purchase Prerequisites: Enterprises need to have or be able to rent an 8×H20-80G level GPU cluster and have corresponding operation and maintenance capabilities. It is recommended to verify the ROI in API mode before deciding whether to switch to self-hosting.

• AI Researchers and Model Developers: Open source weights and a complete fine-tuning/RL pipeline make Hy3 an excellent base model for research and secondary development. 192 Experts The MoE architecture itself also has research value. Unfit Boundary: The hardware requirements for experimental reproduction are high (full weight is about 600GB storage + multi-card GPU).

Use with caution: In real-time interaction scenarios that are extremely sensitive to response delays (such as real-time responses from online customer service), Hy3's TTFT of about 2.65 seconds may not be ideal. It is recommended to choose a small model or distilled version that has been specially optimized for delay. For processing extremely long documents (close to the 256K limit), it is recommended to submit in sections instead of inputting them all at once to avoid the risk of context window overflow.

Summary and Outlook

Hy3 represents an important turning point in China's open source route for large models - Tencent released an open source MoE model under the Apache 2.0 license that surpasses most closed source models on SWE-bench without any geographical restrictions. This is not only a response to Western open source routes such as Meta LLaMA and Mistral, but also a strategic choice: to establish ecological influence through open source rather than relying on API revenue.

Current core advantages: MoE architecture with only 21B activated out of 295B parameters brings excellent cost performance; SWE-bench Verified 78% and GPQA Diamond 90.4% performance are in the first echelon among open source models; Apache 2.0 license completely eliminates legal barriers to commercial use; engineering reliability verified by 50+ Tencent internal products.

Main current limitations: Ecological maturity is lower than DeepSeek and Qwen - third-party tool chains, community tutorials and Chinese technical articles are still being accumulated; 1M ultra-large context is not yet supported; English non-code scenarios are not as sophisticated as pure English models; although the deployment hardware threshold is lower than the same intelligence-dense model, 8×H20-80G is still not easily affordable by small and medium-sized teams.

Follow-up observation points: Whether Tencent will expand the Hy3 series into multi-modal (visual, voice) fields; whether the community will form a rich enough fine-tuning variant and dedicated tool chain around Hy3; whether Tencent will launch an official managed API to simplify the enterprise access process; whether the next version (Hy3.5 or Hy4) will support a larger context window and a higher activation parameter budget.

Procurement and Adoption Risk Assessment: For individual developers and entrepreneurial teams, you can experience the core capabilities of Hy3 with zero risk by paying as you go through third-party APIs such as OpenRouter. It is recommended to incorporate it into your daily coding tool chain for A/B testing. For enterprises planning privatized deployment, it is recommended to deploy trials in non-critical processes (such as internal code review, automated test generation) to verify model output quality and team acceptance before expanding to production-level Agent tasks. Before purchasing, it is important to confirm: whether the GPU model and quantity of the deployed hardware can meet the target load (refer to the official performance data of vLLM/SGLang), the backward compatibility guarantee of model version updates (there is currently no official commitment), and whether the Apache 2.0 license fully covers the expected usage of the enterprise (such as fine-tuned model redistribution). In compliance-sensitive industries (finance, healthcare, government affairs), it is recommended to complete a complete round of red team testing and auditing in a privatized environment before going into production.

Related tools: CrewAI, langchain

Version Info

• Hy3 ：The officially released version is based on large-scale post-training optimization after collecting 50+ product feedback from the Preview version, which significantly improves Agent capabilities, tool calling stability and anti-hallucination capabilities. （2026-07-06）
• Hy3 Preview ：Preview version, the first batch of models trained on the new infrastructure, verifies the feasibility of the 295B MoE architecture, SWE-bench Verified 74.4%. （2026-04-23）
• HY2 (hu energy 2) ：The second-generation hybrid model, SWE-bench is about 53.0%, providing a baseline for the architecture upgrade of Hy3. There is no official precise date yet. （~2025-12）