Tencent Hy3 on Mac: Why 21B Active Does Not Mean 21B of Weights

Newer version: This article covers Tencent Hy3. The newer Tencent Hy4 preview is now available.

Checked October 9, 2026. Tencent Hy3 is an open-weight 295B MoE model with 21B active parameters and Apache-2.0 licensing. The paid OpenRouter slug is tencent/hy3, and tencent/hy3:free is listed as a free, rate-limited variant. 21B active parameters do not mean that a local Mac needs to store only 21B weights. The entire expert pool makes this a server-sized model for most Mac buyers. Tencent model card · OpenRouter.

What Mac users need to know

Tencent Hy3 is an open 295B-parameter Mixture-of-Experts model for reasoning, coding, agents and long productivity workflows. On OpenRouter, you use it as tencent/hy3; the actual open-weight model is available as tencent/Hy3 on Hugging Face. The most important point for Mac users is simple: Hy3 is open and released under Apache 2.0, but it is not a normal local Mac model. Tencent recommends serving it on 8 GPUs such as H20-3e or comparable large-memory GPUs. On a Mac mini, MacBook or normal Mac Studio, Hy3 is realistically an OpenRouter, cloud API or server workflow — not something you run like a 7B or 14B Ollama model.

What is Tencent Hy3?

Hy3 is a large Mixture-of-Experts model from the Tencent Hy Team. The final release follows Hy3 Preview, which Tencent introduced in late April 2026. According to the model card, Tencent collected feedback from more than 50 products, scaled up the post-training pipeline and fixed several real-world product and agent workflow issues.

The focus is not generic chatbot copy. Hy3 is positioned for tasks where the model has to stay reliable across many steps:

• coding agents and repository work
• tool calling and structured outputs
• document processing
• financial modeling
• office and productivity tasks
• frontend design
• game development
• multi-step agent runs
• long contexts with changing constraints

Tencent presents Hy3 as a reliable and cost-effective model option for real productivity tasks. Hy3 is best understood as an open MoE model aimed at agents and long workflows; that positioning does not establish a blanket comparison with individual closed models.

Tencent Hy3 specifications

The “262K” value on OpenRouter and “256K” on the model card are not really a contradiction: 256K commonly means 262,144 tokens. Keep each value attributed to the source that lists it.

Mac memory: activation versus weight storage

At 295 billion total parameters, the mathematical weight-only storage is about 590 GB for BF16 (two bytes per parameter), 295 GB for FP8 (one byte), and at least 147.5 GB for idealized four-bit weights. These are calculated lower-bound payload sizes, not a benchmark or a claim that such a model runs on a particular Mac. Real inference needs additional quantization metadata, KV cache, runtime allocations and operating-system headroom; offloading changes the memory distribution.

So “21B active means it fits like a 21B model” is false. Practical local deployment requires a tested model conversion, compatible runtime and quantization, measured peak memory, and a workload-specific context length. Tencent Hy3 weights.

Why Hy3 is not simply a local Mac model

Hy3 is open, but huge. This is where AI articles often go wrong: open weights does not automatically mean MacBook-friendly.

The Hugging Face card includes Docker Model Runner and quantization links. But for actual serving, Tencent points to vLLM or SGLang and says Hy3 has 295B parameters in total; to serve it on 8 GPUs, Tencent recommends H20-3e or other GPUs with large memory capacity.

For Mac use, that means:

• Hy3 does not simply run locally with Ollama on a Mac.
• Having 21B active parameters does not make a Mac mini sufficient by itself.
• GGUF or MLX support has to be checked for a specific runtime and quantization.
• A blanket recommendation of 16, 24, 32 or 64 GB unified memory would not be reliable.

For Mac users, the practical framing is:

Hy3 is an open server/cloud-class model. On Mac, it makes sense as an API model through OpenRouter or a hosted GPU setup. For private offline workflows, smaller Qwen, Gemma, Mistral or Llama models remain more realistic.

For the local side of that trade-off, start with how much RAM a local LLM needs on a Mac and the Ollama setup guide for the Mac mini M4.

OpenRouter pricing: paid, preview, and free endpoints

OpenRouter’s October 9, 2026 model listing shows $0.132 per million input tokens, $0.528 per million output tokens, and $0.033 per million cached-input tokens for the standard tencent/hy3 offer. These are USD API list rates, not subscription fees. Upstream provider prices, throughput, routing, and availability can vary. Pinning one provider means verifying that provider’s rate and data terms. OpenRouter pricing.

OpenRouter model	Input per 1M	Output per 1M	Scope
tencent/hy3 · listed rate	$0.132	$0.528	paid model routing
tencent/hy3-preview	$0.180	$0.600	older preview version
tencent/hy3:free	$0	$0	rate limited, no guaranteed capacity

U.S. billing: Prices are before any applicable state or local sales tax or provider surcharges. A U.S. billing address does not necessarily guarantee U.S.-only inference; review where OpenRouter’s upstream provider processes and retains your prompts. API pricing is independent of subscription plans.

Free endpoint: OpenRouter lists this as a no-token-charge endpoint but explicitly rate-limits requests. Do not treat it as a paid SLA, stable production capacity, or proof of a particular data-retention policy. OpenRouter free endpoint.

Calculate Hy3’s actual OpenRouter bill: cached input is a subset

OpenRouter lists $0.132 per million standard input tokens, $0.528 per million output tokens and $0.033 per million cache-read tokens for Tencent Hy3, checked October 9, 2026. The discounted cache rate only matters when the selected upstream endpoint actually supports and bills cached reads. Provider-specific rates can differ. Confirm your endpoint and actual usage rather than assume every request gets a cache hit.

Worked examples, not live API measurements; U.S. list USD excluding sales taxes or other surcharges:

Hypothetical request at listed rates	Total input	Cached subset	Output	Calculated price
No cache hit	100,000 tokens	0	10,000 tokens	$0.01848
60% of input receives cache-read price	100,000 tokens	60,000 tokens	10,000 tokens	$0.01254
Longer request without caching	200,000 tokens	0	50,000 tokens	$0.05280

In the cached case, 40,000 input tokens are uncached: 40,000 × $0.132 / 1M + 60,000 × $0.033 / 1M + 10,000 × $0.528 / 1M = $0.01254. The 60,000 cached tokens are already part of the input, so billing them again at the standard input rate would double-count. With no confirmed cache hit, use the first scenario. Tool-heavy calls, reasoning, longer answers and model-provider changes affect real invoices.

Cost and privacy are independent: A lower-rate upstream can have different logging or residency practices. A U.S. card or billing address does not ensure U.S.-only processing. Review the actual host and contract before processing sensitive work.

API setup on Mac

OpenRouter is OpenAI-compatible. On Mac, many OpenAI-compatible clients can keep using the same SDK while changing only the base URL, API key and model name.

read -rs OPENROUTER_API_KEY
export OPENROUTER_API_KEY

import os
from openai import OpenAI

client = OpenAI(
    base_url="https://openrouter.ai/api/v1",
    api_key=os.environ["OPENROUTER_API_KEY"],
)

response = client.chat.completions.create(
    model="tencent/hy3",
    messages=[
        {
            "role": "system",
            "content": "You are a precise Mac and coding assistant. Keep answers concise, flag uncertainty and do not invent benchmarks."
        },
        {
            "role": "user",
            "content": "Explain when I should use a local Mac model and when Tencent Hy3 through OpenRouter makes more sense."
        }
    ],
    temperature=0.7,
)

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

Important: this code uses Hy3 through OpenRouter. It does not download the model to your Mac and it does not run inference on Apple Silicon.

Secure OpenRouter routing: ZDR, provider fallback, and residency

Default routing is not a privacy guarantee. OpenRouter documents automatic upstream selection and fallbacks enabled by default. Agent clients carrying private code or documents should explicitly configure provider routing. This Mac example calls the hosted OpenRouter API, not a local copy of Hy3:

# Read the API key without putting it into shell command history.
read -rs OPENROUTER_API_KEY
export OPENROUTER_API_KEY
curl --fail-with-body -sS https://openrouter.ai/api/v1/chat/completions \
  -H "Authorization: Bearer $OPENROUTER_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{"model":"tencent/hy3","messages":[{"role":"user","content":"Reply with OK only."}],"max_tokens":64,"provider":{"data_collection":"deny","zdr":true,"require_parameters":true,"allow_fallbacks":false}}'

What the documented controls actually do:

• data_collection: "deny" excludes upstream endpoints that OpenRouter identifies as retaining data beyond transient processing or using it for training. It does not replace a review of contracts, subprocessors and actual data transfers.
• zdr: true requires an endpoint classified by OpenRouter as supporting Zero Data Retention. It is not a U.S.-region selection and does not establish full privacy-law compliance.
• require_parameters: true only allows providers that support the requested parameters; allow_fallbacks: false disables automatic alternative-provider failover. To restrict providers further, populate provider.only with exact current provider slugs from the model’s endpoints page. Without an explicit endpoint and regional contract, neither the U.S. nor EU processing location is pinned.
• Failure is expected if the policy cannot be met. If no Hy3 endpoint satisfies the filters, the request can return an error. Do not silently remove safeguards simply to get a completion.

The example uses a synthetic prompt only. We have not executed a live test of Hy3’s compatibility with ZDR endpoints, nor verified regional availability for this filter combination. U.S. teams handling regulated data must assess contracts and real provider processing locations independently; filtering is not a blanket compliance certification.

Reasoning modes: do not blindly use the deepest mode

Hy3 supports configurable reasoning according to OpenRouter and the model card. The model card shows reasoning_effort through chat_template_kwargs for local OpenAI-compatible servers:

• no_think: direct answer, default in the example configuration
• low: lighter reasoning mode
• high: deeper reasoning for math, coding and complex multi-step tasks

Do not publish full chain-of-thought traces. What matters is the outcome: better planning, more stable tool use and less task drift. Reasoning modes are latency, cost and quality controls, not magic quality switches.

What improved over Hy3 Preview?

The final Hy3 release is not just a renamed preview. Tencent names several improvements from product feedback and post-training:

These are vendor claims, not independent measurements. They are useful signals, but they do not replace your own tests on real workflows.

Benchmarks: strong, but read them carefully

The Hugging Face page shows several evaluation results for Hy3, including:

These numbers are useful context, but they do not establish a general model ranking. Benchmarks depend heavily on the harness, scaffold, tool access, prompting, reasoning mode, timeout and scoring logic. That is especially true for SWE-bench-style evaluations. They prove neither that Hy3 beats Claude nor that it is the best coding model or better than every local Mac model.

For real Mac workflows, the deciding question is whether Hy3 is reliable, affordable and acceptable from a data-handling perspective in your own tooling.

When Hy3 makes sense for Mac users

Hy3 can be useful if you work on a Mac but consciously accept cloud or server inference.

Good use cases:

• reviewing a large repository
• planning a multi-file refactor
• structuring an agent run with tools
• analyzing data or spreadsheet logic
• summarizing technical documentation
• keeping long chat or project context stable
• designing frontend components
• testing browser or research agents

Poor use cases:

• private documents without cloud processing
• offline work on a train or at university
• small everyday questions where a local 4B to 14B model is enough
• sensitive customer code without clear data approval
• tasks where latency matters more than model strength

Hy3 vs local Mac models

The best workflow is hybrid: local models for private, fast and offline tasks; Hy3 for large agentic tasks where cloud processing is acceptable.

Hy3 vs Hy3 Preview

Hy3 Preview was the April 2026 preview. OpenRouter still lists it separately as tencent/hy3-preview. For new integrations, use the final tencent/hy3 release.

The license difference matters: do not infer the final Hy3 license from old preview information. The final Hy3 model card lists Apache 2.0.

Privacy in the United States: OpenRouter is not the final model host

OpenRouter brokers access to an upstream inference provider. Prompts and tool outputs leave the Mac, and a U.S. account or English prompt does not by itself ensure U.S.-only processing, retention, or hosting. Before uploading customer code, unpublished work, or personal information, review provider filtering, logging, model-training policy and contractual data terms. For work that must stay offline, use a model you have verified runs entirely on-device.

Mac deployment choice

Tencent Hy3 is an open MoE model with 295B parameters, 21B active, 256K context, Apache 2.0 licensing and OpenRouter access through tencent/hy3. Its positioning is focused on agent workflows.

For Mac users, the crucial message is still: Hy3 is not suddenly a local Mac model. It is an open but server-sized model. On Mac, it makes sense as a cloud/API complement to local models. If you want privacy, offline work and Apple-Silicon inference, use smaller local models. If you need long agent runs, broader context understanding and strong tool workflows, test Hy3 through OpenRouter or dedicated GPU infrastructure.

Sources

• Tencent Hy3 on Hugging Face: https://huggingface.co/tencent/Hy3
• Tencent Hy3 on OpenRouter: https://openrouter.ai/tencent/hy3
• OpenRouter Endpoint-Preise: https://openrouter.ai/tencent/hy3-20260706/
• Tencent Hy3 Preview on OpenRouter: https://openrouter.ai/tencent/hy3-preview
• Tencent Hy3 Free on OpenRouter: https://openrouter.ai/tencent/hy3:free
• OpenRouter Models API: https://openrouter.ai/api/v1/models
• Tencent Hy3 Preview on Hugging Face: https://huggingface.co/tencent/Hy3-preview
• Tencent News on Hy3 Preview: https://www.tencent.com/en-us/articles/2202320.html