Tencent Hy4: API Pricing, Computer Use and Mac Limits

Researched and edited with AI assistance. Editorial responsibility: Julian Dominic Altmann. Facts checked November 4, 2026. Source analysis, not an independently executed API or Mac benchmark.

Tencent Hy4 launched on November 4, 2026 with support for image understanding and computer-use workflows. Tencent also claims improvements over its August Preview in long-horizon agents, frontend development and workplace tasks. The model is available through Tencent Cloud TokenHub and OpenRouter.

There is an important distinction for developers: API availability is not the same as open-weight availability. Tencent says the weights and system cards for the final release are coming, but they were not verified as public downloads at the time of this review. The existing Hy4 Preview checkpoint is a separate release. For ordinary Mac hardware, hosted inference is the realistic option. Tencent announcement

Tencent Hy4 specifications: what is actually documented?

Property	Verified November 4 status
Developer	Tencent Hy Team
Release	November 4, 2026
Tencent Cloud model ID	hy4
OpenRouter model ID	tencent/hy4
Input / output	Text and images / text
Context length	1,048,576 tokens listed by OpenRouter
Tencent Cloud maximum input	960k tokens
Maximum completion	65,536 tokens on OpenRouter; 64k on Tencent Cloud
API features	Reasoning, tool calls, structured output, cache
Final weights	Announced, not yet confirmed publicly downloadable
OpenRouter promotional prices	$0.417 input / $1.251 output per million tokens
Tencent Cloud sale	50% off until Dec. 31, 2026, 23:59:59 Beijing time

A model’s nominal context window, the maximum prompt size and the maximum completion allowance should not be treated as the same limit. Route-specific ceilings still matter. Sources: Tencent model list and OpenRouter’s Hy4 listing.

Hy4 versus Hy4 Preview: the meaningful differences

The August 28 Preview introduced a large Mixture-of-Experts architecture. Its official model card reports 770 billion total parameters and 49 billion active parameters per token. November’s Hy4 introduces vision and extends Tencent’s computer-use ambitions into GUI and command-line workflows.

Capability	August Preview	November Hy4
Image input	Not documented as an available capability	Available
Computer use	Not promoted as a new capability	Emphasized
Tool calling	Supported	Supported
Context class	1M	1M
Official checkpoint	Downloadable Preview	Final weights still announced
Published weight license	Apache 2.0 for Preview	Final license not yet verified

Do not copy the 770B/49B number onto the final release as an independently verified November architecture specification. It is confirmed for the Preview checkpoint. Tencent has not yet made a corresponding final model card available in the material reviewed here.

Why vision and computer use matter for agents

A screenshot of a broken layout can now be included in a prompt alongside repository files. The visual step can help identify missing UI states, chart interpretation errors or design mismatches. Tencent also describes workflows that operate desktop applications such as Blender and CAD software.

Those workflows require an agent environment with explicitly granted tool access. A standalone API call cannot automatically operate a user’s Mac, write code to disk or run a browser. Tencent mentions platforms such as Hermes and OpenCode as routes to the model, but compatibility and long-run task completion must still be evaluated in each integration.

Hy4 API pricing in U.S. dollars

OpenRouter currently advertises the following promotional dollar rates per one million tokens:

Billing category	Promotional rate
Uncached input	$0.417
Output	$1.251
Cached input / cache read	$0.021

Its standard crossed-out rates are $0.834 input, $2.501 output and $0.042 cache read per million tokens. These are OpenRouter-specific figures, not a universal commitment from every upstream API vendor.

Tencent Cloud separately documents CNY 6 / 18 / 0.30 per million tokens for uncached input, output and cache hits. A 50% promotion through December 31, 2026 at 23:59:59 Beijing time means a calculated effective price of CNY 3 / 9 / 0.15 during the eligible period. Currency, tax and endpoint conditions remain provider-specific. Official Tencent pricing

For U.S. readers: Dollar pricing does not imply U.S.-only processing. Taxes may be additional depending on the account, jurisdiction and billing provider. Tencent’s local-currency listing is not a U.S. regional service-level commitment.

Three original cost scenarios

Assume no cache hits, taxes, tool charges or retry overhead. These are calculations rather than measured application usage.

Workload	Regular input	Output	API estimate
Occasional testing	1M	0.2M	$0.67
Frequent development	10M	2M	$6.67
Heavy agent usage	100M	20M	$66.72

The formula is cost = input millions × $0.417 + output millions × $1.251. A separate scenario with 10M new input tokens, 20M cache-read tokens and 2M output tokens would be $7.09, if all repeated tokens qualify for the quoted cache rate.

Pricing per token is only one component of a coding agent’s operating cost. Failed verification, retries, tool calls and project-context replays can dominate the eventual cost per accepted task.

How to call Hy4 from an Apple Silicon Mac

Mac developers can access the model through the OpenAI Python SDK without installing the model weights. Tencent’s officially documented base URL is https://tokenhub.tencentmaas.com/v1 and its model ID is hy4.

import os
from openai import OpenAI

client = OpenAI(
    api_key=os.environ["TENCENT_API_KEY"],
    base_url="https://tokenhub.tencentmaas.com/v1",
)

result = client.chat.completions.create(
    model="hy4",
    messages=[{
        "role": "user",
        "content": "Identify the likely root cause of this Python test failure."
    }],
    temperature=0.9,
)
print(result.choices[0].message.content)

The example requires an enabled Tencent TokenHub account, valid API credentials and the openai package. Put the key in an environment variable and never commit it to Git. The request follows Tencent’s official guide; we did not execute it against live credentials for this article.

When using OpenRouter instead, configure its own endpoint and credentials with model ID tencent/hy4. Mixing provider-specific IDs is a common configuration mistake.

Why Hy4 cannot realistically run locally on an ordinary Mac

The public Preview repositories are enormous: Hugging Face lists approximately 1.56 TB for the standard checkpoint and 814 GB for the FP8 variant. These are file sizes, not a direct measurement of runtime RAM, but they explain why a regular 32GB or 64GB Apple Silicon Mac cannot hold the complete released Preview checkpoints in unified memory. Preview files · FP8 files

As a lower-bound thought experiment, a hypothetical four-bit representation of 770 billion Preview parameters takes 385 GB in raw weights alone (770 billion × half a byte). Real quantization has additional metadata and inference also needs KV cache and temporary buffers. This is not a published Hy4 four-bit file or a benchmark.

Tencent’s publicly available Preview serving guidance is aimed at multi-GPU environments with frameworks such as vLLM and SGLang. If fully local data processing is required, smaller open-weight models are a more practical Mac route.

Benchmarks: promising launch claims, not independent rankings

Tencent and Tencent Financial Technology organized a blind comparison with 213 finance professionals evaluating 579 real-world tasks. Tencent reported these scores:

Model	Tencent-reported score
Hy4	3.04
DeepSeek V4.1 Flash 1017	2.97
Kimi K3.1 Pro	2.68
GLM-5.3	2.68

Independent practitioners participated, but the evaluation was organized by Tencent and partners. The results are not a universal model leaderboard or proof of better coding-agent performance.

Tencent also reports a reduction in its internally measured hallucination rate from 12.95% for Preview to 7.2% for Hy4: a 5.75-percentage-point drop, or about 44.4% relative. This does not establish accuracy for a different benchmark, coding harness or deployment setup. Tencent release

A more useful developer benchmark is a repeated, version-controlled repository task with the same tools and success criteria: did the patch fix the bug, did the tests pass, how many iterations were needed, and what did the run cost?

Hosting geography, final weights and Hy4 Flash

As of November 4, 2026, OpenRouter states Hy4 is not available via EU in-region routing. That is important for international teams with strict EU-only processing policies; a U.S. account also does not prove U.S. data residency. Verify upstream providers, data retention, logging and contractual commitments before sharing proprietary code or regulated data. OpenRouter routing

Tencent plans to publish the final weights and system cards. It also mentions a lower-cost Hy4 Flash, whose precise availability and pricing were not established from the reviewed official documentation. An English press-release subheading is ambiguous, but the surrounding text describes Flash as scheduled to arrive later.

Hy4 is an inexpensive API candidate for screenshot-aware development and longer agent tasks during the promotion. It is not a realistically downloadable full model for an ordinary Mac today, and vendor benchmark scores still need reproducible third-party validation.

For the previously released checkpoint, see Hy4 Preview: architecture, API limits and Apple Silicon constraints.