Tencent Hy3: Free API Access to a 295B Open Model

Tencent Hy3: Free API Access to a 295B Open Model

Tencent released Hy3 on July 6, 2026: a 295-billion-parameter mixture-of-experts model under the Apache 2.0 license, with free API access on OpenRouter through July 21, 2026. If you have an OpenRouter account, you can call it right now at zero cost for two weeks.

This post covers what Hy3 actually is, the verified benchmark numbers, how to use the free API window, and what Apache 2.0 means if you want to build products on top of it.

What Hy3 Is

Hy3 is a sparse MoE (mixture-of-experts) model. The headline 295B refers to total parameters across all expert layers. At inference time, only a subset of those parameters activates per token: 21 billion active parameters with 192 experts and top-8 routing (the top 8 of 192 experts process each token). This keeps compute closer to a 21B-scale model while retaining the capacity of a much larger dense network.

Context window: 256K tokens (per the HuggingFace model card; OpenRouter reports 262K due to rounding differences; the model card is the primary source).

Verified Specs

All figures below are sourced from the Tencent Hy3 HuggingFace model card unless noted otherwise.

Benchmark Results (Vendor-Reported)

The figures below are from the official Hy3 model card and research page. They are Tencent's own reported numbers and have not been independently replicated by this site.

Tencent claims Hy3 "rivals trillion-scale flagship models." That is the vendor's positioning, not an independently verified claim.

Comparison to Frontier Models

For SWE-bench Pro, here is where Hy3 sits against frontier models (all figures are vendor-reported as of July 2026):

On SWE-bench Pro, Hy3 scores below the current Claude frontier. On GPQA Diamond (90.4), Hy3 performs at or above publicly reported scores for most 2025-era flagship models, though direct comparisons require running the same eval harness on the same eval set.

How to Use the Free API Window

The free tier is available through July 21, 2026 via the OpenRouter endpoint tencent/hy3:free. You need an OpenRouter account but no payment method during the free window.

Python

from openai import OpenAI

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

response = client.chat.completions.create(
    model="tencent/hy3:free",
    messages=[
        {"role": "user", "content": "Explain mixture-of-experts in one paragraph."}
    ],
    max_tokens=512,
)

print(response.choices[0].message.content)

JavaScript / Node.js

import OpenAI from "openai";

const client = new OpenAI({
  baseURL: "https://openrouter.ai/api/v1",
  apiKey: process.env.OPENROUTER_API_KEY,
});

const response = await client.chat.completions.create({
  model: "tencent/hy3:free",
  messages: [{ role: "user", content: "Hello" }],
});

console.log(response.choices[0].message.content);

The :free suffix is the key. After July 21, the model switches to paid pricing. Plan your evaluation work before that date if you want zero-cost testing.

See our free models hub for other models with free API tiers.

Agentic Use: What Tencent Claims vs. What to Expect

Tencent positions Hy3 as an agentic model, pointing to its SWE-bench Verified score (78.0) and specialized coding benchmarks.

The 256K context window is large enough for multi-file repositories and long tool-call chains without chunking. The MoE architecture means inference can be efficient relative to parameter count when running on appropriate hardware (not all cloud providers support MoE natively, so check your serving stack before committing to it in production).

Practitioner caveats:

• Benchmark numbers are measured under controlled conditions. Real agentic performance depends heavily on system prompt design, tool definitions, and retry logic.
• The free window is good for evaluation, not production. Rate limits apply during the free period. Check OpenRouter's docs for current limits.
• 295B MoE models require significant GPU memory to self-host. If you plan to run your own inference after the free window, budget accordingly (multi-GPU setup or a dedicated cloud endpoint).
• Deep SWE (28.0) and Apex Agents (25.6) are lower relative to SWE-bench Verified (78.0). These harder agentic benchmarks reflect what the model does under real scaffolding pressure. Weight them accordingly when planning agent deployments.

For coding-focused free alternatives, see Free AI Models for Coding.

Apache 2.0: What It Means for Commercial Use

Hy3 is released under the Apache 2.0 license. Key implications:

• You can use the weights and outputs commercially with no royalty or revenue share.
• You can fine-tune and redistribute modified versions.
• You must preserve the Apache 2.0 notice and attributions in distributions.
• There is no "non-commercial only" restriction common to some open-weight models.

Apache 2.0 is one of the most permissive licenses available for ML models. It puts Hy3 in the same tier as Llama 3 and Mistral models for commercial viability. If Apache 2.0 commercial use is a requirement for your project, Hy3 qualifies.

See our Free AI Models for Commercial Use post for a broader comparison of open-model licenses.

n8n and Workflow Integration

OpenRouter-compatible models work with any n8n HTTP Request node or the OpenAI-compatible node by swapping the base URL to https://openrouter.ai/api/v1 and setting the model ID to tencent/hy3:free. No custom plugin required.

For n8n-specific free model setup, see Free AI Models for n8n.

FAQ

Is Hy3 actually free? Yes, through July 21, 2026, via the OpenRouter tencent/hy3:free endpoint. After that date, standard OpenRouter pricing applies. Tencent has not announced post-window pricing publicly as of this writing (July 7, 2026).

Can I self-host Hy3? The weights are public on HuggingFace under Apache 2.0. You can self-host, but a 295B MoE model requires significant GPU infrastructure. Plan on multi-GPU nodes or a dedicated cloud inference setup.

How does Hy3 compare to Llama and Mistral? Llama 3 and Mistral models are also Apache 2.0 but at smaller scales. For benchmark comparisons, refer to vendor-reported evals for each specific version. Direct comparison requires running the same eval harness and eval set, which we have not done here.

What is the context window? 256K tokens per the HuggingFace model card. OpenRouter displays 262K due to rounding convention differences between providers.

Does Hy3 support function calling? The model card lists tool-use benchmarks (HLE with tools: 53.2), indicating the model was trained for tool use. OpenRouter exposes it through the standard chat completions API with tools in the OpenAI format. Verify tool-calling behavior in your own evals before deploying to production.

When does the free window end? July 21, 2026. After that date, you will need OpenRouter credits to continue using the model.

---

Sources: Tencent Hy3 HuggingFace model card (huggingface.co/tencent/Hy3), Tencent Hunyuan research page (hy.tencent.com/research/hy3), OpenRouter model page. All benchmark figures are vendor-reported. Claude benchmark figures are Anthropic-reported.