GenGrades

Tencent Hy4 review: price, benchmarks and open weights

Published Updated

TencentTencent Hy4 logoAPI availableClosed weightsLicense: Weights not released yet; Hy4 preview is Apache 2.0

What Tencent Hy4 is

In one or two sentences, before anything else.

Tencent Hy4 is Tencent's flagship language model, released November 4, 2026, and its first Hy model with vision and computer use. Until December 31 Tencent bills half its 6-yuan input and 18-yuan output list price per million tokens; OpenRouter charges 0.417 and 1.25 USD. Mercor ranks it third on APEX-Agents, but its weights are unreleased and coding is slow.

Main features

What it can do, one sentence each, and whether we actually tested it.

  • Image input

    Reads images in the same message as text, for charts, screenshots, scanned documents and photos: PNG, JPEG, WebP or still GIF, up to 10 MB each and 600 per request including the conversation history.

    Per Tencent Cloud's Hy call guide, images over 5,500 by 5,500 pixels are rejected and detail beyond 2,048 by 2,048 is not seen. Hy4 preview read text only.

  • Computer use

    Operates software through both the command line and the graphical interface; Tencent trained it on Blender, CAD tools and long tasks that cross several apps.

    Per Tencent's release page. Tencent Cloud's API guide covers image input and tool calls but documents no separate computer-use tool; the press runs so far, by InfoQ and 科创板日报, used Tencent's WorkBuddy agent.

  • 1M-token context

    Takes a context window of 1,024K tokens, with up to 960K tokens of input and 64K of output per request.

    Per Tencent Cloud's TokenHub model list; OpenRouter lists 1,048,576 tokens of context and 65,536 of output.

  • Three reasoning levels

    Thinks step by step by default (reasoning_effort high) and can be set to low or switched off with none.

    AI/ML API, which resells Hy4, found low passed all 8 of its coding tasks with 22% fewer reasoning tokens, while none failed 3 of 10 runs.

  • Agent work across office apps

    Carries long tasks across documents, spreadsheets, slides, email and chat; Tencent's WorkBuddy agent uses it to co-edit Word, Excel and PowerPoint files with the user.

    On Mercor's APEX-Agents leaderboard, Hy4 at high effort completes 77.6% of professional-services tasks on the first try, third behind Gemini 4 Carbon and Gemini 4 Argon.

  • Finance analysis

    Was trained on thousands of finance workspaces and more than ten thousand tasks covering fund research, securities analysis, quant strategies, asset allocation, corporate finance and tax.

    In Tencent's own crowd test, 213 experts graded 579 tasks and gave Hy4 3.04 out of 4, against 2.97 for DeepSeek-V4.1-Flash-1017.

  • Game and front-end building

    Is tuned for Unreal Engine, Unity and Godot projects, early multiplayer game code, and front-end state, logic and back-end wiring.

    Tencent reports gains of 19.1% on GameDev and 9.8% on GameLogic over the preview; InfoQ had it build and publish a tower-defence web game in three rounds totalling about 55 minutes.

  • OpenAI- and Anthropic-style API

    Serves the OpenAI Chat Completions, OpenAI Responses and Anthropic Messages formats, with function calling, structured output and prompt caching, so agent tools connect by changing the base URL.

    Per Tencent Cloud's Hy call guide; 科创板日报 reports the API also reached OpenRouter, Hermes and OpenCode at launch.

Who it is for

The people it helps and the problems it solves, and where it ranks.

Built for

  • Teams building office and finance agents that read documents, spreadsheets and screenshots
  • Developers who want a 1M-token model with image input at 0.417 USD in and 1.25 USD out per million tokens on OpenRouter
  • Companies already on Tencent Cloud, WorkBuddy, CodeBuddy or WeCom, where Hy4 is already built in
  • Game studios prototyping in Unity, Unreal Engine or Godot
  • Self-hosters with an 8-GPU B300 or MI355X node who can run Hy4 preview's Apache 2.0 weights now

Problems it solves

  • Agents that stall on long, multi-app professional tasks such as building a valuation model from filings
  • Text-only models, Hy4 preview and GLM-5.3 among them, that cannot read the chart, scan or screenshot in a task
  • Per-task cost of long agent runs: Vals measured 1.72 USD per Vals Index task for Hy4 against 8.52 USD for GLM 5.3
  • Models that invent an answer when the information is missing

What it is good at

Public tests other people ran: what they gave it and how it did.

  1. Case 1

    A tower-defence web game built and shipped in three rounds

    Input

    InfoQ's test in Tencent's WorkBuddy agent on November 4, 2026: "make a game, from zero to ready to publish", then two follow-ups asking for enemy types, tower evolution and difficulty levels, then multiple maps and a leaderboard.

    What happened

    Hy4 finished the three rounds in about 55 minutes (14, 18 and 23 minutes), writing simulation scripts to balance difficulty and checking the page in Chrome before each release. InfoQ found the game smooth early on, but later waves turned into idle clicking: the simulations proved the game could be won, not that it was fun.

  2. Case 2

    A comparative finance report from company filings

    Input

    A 科创板日报 reporter's request in WorkBuddy on November 4, 2026: download the quarterly, half-year and annual reports of Tencent, Alibaba and Baidu since 2020 and compare them from four analyst viewpoints.

    What happened

    The run took 41 minutes 51 seconds and returned a 12-page PDF with 8 charts, an Excel workbook with 183 formulas that can be recalculated, and a download log with each URL and failure reason. The reporter judged it close to an intern's first draft.

  3. Case 3

    Checking an order form against a goods-received note from one image

    Input

    InfoQ's test on November 4, 2026: an image with an order on the left and a receipt record on the right, with instructions to list mismatches, quote both sides and separate recorded differences from proven errors.

    What happened

    Hy4 matched the lines correctly, found that only two items agreed fully and noted the inspection status was still pending. It also miscounted the remaining items as two instead of three and claimed the record was filled in by hand without evidence in the image.

  4. Case 4

    Scheduling two robots on a grid map

    Input

    InfoQ's visual reasoning test on November 4, 2026: a warehouse map image with one passing bay, two robots with opposite goals, and a request for collision-free routes and the earliest finish.

    What happened

    Hy4 produced a schedule with no shared cell or swapped edge, showed the passing bay was required, and proved that eight time steps is the minimum, which its schedule met.

  5. Case 5

    Coding tasks at three reasoning levels through an API

    Input

    AI/ML API, which sells Hy4, ran coding, Three.js, tool-calling, long-log and image tasks on November 4, 2026, 23 runs at the default setting plus runs at low and none.

    What happened

    All 23 default runs passed, but a coding task took a median 4.4 minutes with about 18,000 reasoning tokens, against 25 seconds for DeepSeek V4.1 Flash. At low, 8 of 8 coding tasks passed for 0.039 USD each; with reasoning off, 3 of 10 failed.

Where it falls short

What it does badly, with the evidence, and what we could not settle.

Weak

Weak

  • The weights are not released.

    Evidence: Tencent's release page says it plans to open-source them with a system card and gives no date. On November 6, 2026 Tencent's Hugging Face organisation offered only Hy4 preview, and Artificial Analysis lists Hy4 as a proprietary model until weights appear.

  • Behind Kimi K3.1 Pro on the independent indexes.

    Evidence: Artificial Analysis gives Hy4 35 on its Intelligence Index against 46.5 for Kimi K3.1 Pro. On the Vals Index of November 5 it scores 59.35%, below Kimi K3.1 Pro (62.52%), DeepSeek V4.1 Flash 1017 (61.26%) and GLM 5.3 (59.73%).

  • Weak at terminal work and hard maths.

    Evidence: In Tencent's own table Hy4 scores 44.4 on Terminal-Bench 4.0 against 65.7 for Kimi K3.1 Pro and 66.4 for Claude Opus 5.5; Artificial Analysis measured 37.4% and Vals 39.4%. Tencent also reports 23.2 on its Goal2Repo test against 74.0 for Opus 5.5, and 13.0 on its hardest maths set against 97.5 for GPT-6.1 Sol.

  • Slow and verbose.

    Evidence: Artificial Analysis counted 160 million output tokens to run its index, against a median of 72.6 million, and 103 seconds to the first answer token. AI/ML API saw a median 4.4 minutes per coding task.

  • Not for anyone who must self-host Hy4 now, needs EU-region routing, or runs terminal-heavy coding agents.

    Evidence: No Hy4 weights are public, OpenRouter serves Hy4 from Tencent alone and says it is not available through EU in-region routing, and Kimi K3.1 Pro and Claude Opus 5.5 lead it by over 20 points on Terminal-Bench 4.0 in Tencent's own table.

Uncertain

Uncertain

  • When the weights arrive, under which licence and at what size.

    Why unsure: Hy4 preview's weights are Apache 2.0 with no territory clause. For Hy4, Tencent names neither a date nor a licence, and its release page gives no parameter count; 科创板日报 and InfoQ put the Hy4 series at 770 billion parameters.

  • Whether it really guesses less.

    Why unsure: Tencent's two figures come from its own tests and disagree (15.76% to 5.45%, or 12.95% to 7.2%). On Artificial Analysis's Omniscience test, which rewards saying "I don't know", Hy4 scored 5.4 against 29.9 for Kimi K3.1 Pro.

  • How well computer use works outside Tencent's own products.

    Why unsure: The 78.9 on OSWorld 2.1 is Tencent's own run, and the press runs so far used Tencent's WorkBuddy agent. No outside lab had published a computer-use run by November 6, 2026.

  • The price after December 31.

    Why unsure: TokenHub bills Hy4 at half its list price until December 31, 2026. OpenRouter marks its price 50% off without an end date. Tencent has announced a cheaper Hy4 flash without a date or price.

Price, limits and formats

The numbers, each with its source and the date we checked it.

SpecValueNote and source
ReleasedNovember 4, 2026; Hy4 preview on August 28, 2026

Source: cloud.tencent.comchecked

Previous generationHy3: 295B total, 21B active, 256K context, text only, Apache 2.0

TokenHub price 1 yuan in, 4 out, 0.25 cache hit per 1M tokens.

Source: huggingface.cochecked

Size770B total, 49B active per token (Hy4 preview's card)

Tencent's Hy4 release page gives no parameter count; 科创板日报 and InfoQ put the Hy4 series at 770B.

Source: huggingface.cochecked

Architecture (preview)Mixture of experts: 256 routed and 1 shared expert, 8 active; gated sparse attention with IndexCache; one 10B MTP layer

Source: huggingface.cochecked

Inputs and outputText and images in, text out

Source: cloud.tencent.comchecked

Context window1,024K tokens; up to 960K input and 64K output per request

Source: cloud.tencent.comchecked

Image limitsPNG, JPEG, WebP or still GIF; 10 MB each; 600 per request; best at 128 to 2,048 pixels a side

Source: cloud.tencent.comchecked

Reasoninghigh (default), low or none

Source: openrouter.aichecked

API formatsOpenAI Chat Completions and Responses; Anthropic Messages

Source: cloud.tencent.comchecked

TokenHub priceList: 6 yuan per 1M input tokens, 18 output, 0.3 cache hit; billed at half (3 / 9 / 0.15) until December 31, 2026

Discount ends 23:59:59 Beijing time; same rates in Guangzhou and Singapore; no peak or off-peak rate for Hy4.

Source: cloud.tencent.comchecked

Free accessDaily free quota in WorkBuddy until November 15, 2026, 23:59; more for subscribers

No night-time window for Hy4; WorkBuddy's 23:00-08:00 free window was for Hy4 preview, set to end October 31.

Source: finance.sina.com.cnchecked

OpenRouter priceNow 0.417 USD in, 1.25 out, 0.021 cached per 1M tokens, marked 50% off 0.834 / 2.501 / 0.042

OpenRouter gives no end date for the discount. One provider, Tencent, serving 4-bit (FP4) weights.

Source: openrouter.aichecked

Open weightsNot yet; Tencent plans to release them with a system card

Source: hunyuan.tencent.comchecked

Preview licenceApache 2.0, with no territory or user-count limits

Source: huggingface.cochecked

Preview weight sizeAbout 1,560 GB in BF16; about 814 GB in the official FP8 build

Summed from the safetensors files listed on Hugging Face.

Source: huggingface.cochecked

Hardware for the preview8 B300 GPUs, 16 B200s or 8 AMD MI355X

Source: recipes.vllm.aichecked

Data useTokenHub says it does not use customer data for training

Source: cloud.tencent.comchecked

Consumer accessYuanbao app (iOS 2.92.0, November 6, 2026), WorkBuddy, CodeBuddy and ima

WorkBuddy, CodeBuddy and ima per Tencent's release page.

Source: apps.apple.comchecked

APEX Agents (Mercor)77.6% pass@1 at high effort, third of all models listed

Source: mercor.comchecked

Vals Index59.35%, 13th overall; 1.72 USD per task

Vals Index version 2.2, updated November 5, 2026.

Source: vals.aichecked

Common questions

Answer first, then why, an example, the limits and what to do next.

What is Tencent Hy4?

Tencent Hy4 is the flagship large language model of Tencent's Hy team, formerly called Hunyuan, released on November 4, 2026. Compared with Hy4 preview from August it adds image input and computer use and was trained further for agent, office, finance and game-development work. It answers in text and runs through Tencent Cloud's TokenHub, OpenRouter and Tencent apps such as Yuanbao and WorkBuddy.

Is Hy4 open source?

Not yet: on November 6, 2026 Hy4 was available only as a hosted model. Tencent says it will release the weights and a system card but has given no date or licence. Hy4 preview, the August version that reads text only, can be downloaded under Apache 2.0.

How much does Hy4 cost?

Until December 31, 2026, Tencent Cloud's TokenHub bills 3 yuan per million input tokens, 9 yuan per million output tokens and 0.15 yuan for cached input, half the list price of 6, 18 and 0.3 yuan that Hy4 shares with the preview. OpenRouter currently charges 0.417 USD in, 1.25 USD out and 0.021 USD cached, which it marks as 50% off 0.834 and 2.501 USD, with no end date.

Is Hy4 free?

Partly, inside Tencent's apps: WorkBuddy gives a small daily free quota for Hy4 until November 15, 2026, 23:59, with more for subscribers, and the free Yuanbao app added Hy4 on November 6. WorkBuddy's night-time free window (23:00 to 08:00) applied to Hy4 preview and was set to end on October 31; neither Tencent's Hy4 announcements nor its price page show a night rate for Hy4. The API is paid per token, and TokenHub's free-trial list does not include Hy4.

What changed from Hy4 preview to Hy4?

Hy4 reads images and can operate a computer; the preview handled text only. Tencent scaled up pre- and post-training, and its own APEX-Agents run rose from 53.9 to 75.2 and Terminal-Bench 4.0 from 26.3 to 44.4. The price and the 1M-token context stayed the same, and the weights, public for the preview, are not out yet.

What is the difference between Hy4 and Hy3?

Hy4 is the larger, costlier model and reads images; Hy3 is the cheap text-only one. Hy3 has 295 billion parameters with 21 billion active and a 256K-token window, and costs 1 yuan in and 4 yuan out per million tokens on TokenHub against Hy4's 6 and 18; Hy4 adds image input, computer use and a 1,024K window. In InfoQ's side-by-side business-plan test Hy4 took about 90 seconds longer but reasoned more deeply. Hy3's weights are public under Apache 2.0; Hy4's are not yet.

Where does Hy4 rank on public leaderboards?

Near the top for agent work and mid-table overall. Mercor's APEX-Agents puts it third of all models, at 77.6% of tasks completed on the first try behind Gemini 4 Carbon and Gemini 4 Argon and above Claude Opus 5.5. The Vals Index of November 5 ranks it 13th (fourth of the 19 models Vals labels open-weight), and Artificial Analysis places it 40th of 69 with a score of 35, below Kimi K3.1 Pro's 46.5.

Is Hy4 better than Kimi K3.1 Pro?

Only on some agent tasks. Hy4 leads on Mercor's APEX-Agents (77.6% against 65.5%) and is cheaper per token, but Kimi K3.1 Pro scores higher on Artificial Analysis's index (46.5 against 35), on the Vals Index and on Terminal-Bench 4.0. Kimi's weights are already on Hugging Face.

Hy4 vs DeepSeek V4.1 Flash: which is better?

DeepSeek V4.1 Flash is the better pick for fast, cheap coding, and Hy4 for long office-style agent work. DeepSeek costs 0.30 USD in and 1.20 USD out at peak, half off peak, has MIT weights and finished an AI/ML API coding task in 25 seconds against Hy4's 4.4 minutes. Hy4 scores higher on APEX-Agents and Artificial Analysis's index; DeepSeek scores higher on the Vals Index.

Can I run Hy4 locally?

Not the final Hy4, whose weights are unreleased. You can run Hy4 preview: its official FP8 weights take about 814 GB, and vLLM's recipe runs it on 8 B300 GPUs, 16 B200s or 8 AMD MI355X accelerators. That is a data-centre node, not a workstation.

Does Tencent train on what I send to Hy4?

No, not through the TokenHub API: Tencent Cloud's TokenHub FAQ says customer data is used only to serve the customer and never for training. Requests via OpenRouter go to Tencent, which OpenRouter lists as headquartered in China with its data centre in Singapore.

Can I use Hy4 in Europe?

Yes through Tencent's API or OpenRouter, but OpenRouter says Hy4 is not available through its EU in-region routing. The preview's Apache 2.0 licence has no territory clause, unlike the earlier Tencent Hunyuan Community License, which excluded the EU, the UK and South Korea.

What the public tests show

This review is built from what could be checked on November 6, 2026: Tencent's Hy4 release page and benchmark appendix, the Hy4 preview model card and licence, Tencent Cloud's TokenHub documentation and price list, launch coverage by IT之家, 科创板日报 and InfoQ, the Mercor APEX-Agents, Vals Index and Artificial Analysis leaderboards, test runs published by InfoQ and by the reseller AI/ML API, and the price pages of five rival models. Hy4 is two days old: Mercor, Vals and Artificial Analysis have ranked it, most other scores are Tencent's own, and each is labelled with who produced it.

What Hy4 is and who makes it

Hy4 is a large language model from the Hy team at Tencent, the Shenzhen company behind Weixin and WeChat (Tencent). The team publishes in Chinese as 腾讯混元, and its earlier models carried the Hunyuan name in English. It is a reasoning model for long tasks: writing and fixing code, building games and web pages, analysing finance documents and working across office apps. It reads text and images and answers in text (TokenHub model list).

Tencent released Hy4 on November 4, 2026, 68 days after Hy4 preview (TokenHub updates). Tencent's senior executive vice president 汤道生 (Dowson Tong) had announced it as coming soon at Tencent's Global Digital Ecosystem Conference on October 29 (IT之家). On launch day Hy4 went live on Tencent Cloud's TokenHub and OpenRouter and in Tencent's WorkBuddy and CodeBuddy agents, the Yuanbao assistant, the ima note app and WeCom's assistant (科创板日报, IT之家); the Yuanbao iPhone app added it on November 6 (App Store).

Tencent is unusually direct about the limits. Its release page calls Hy4 one of the strongest open models but says a large gap to the top closed models remains, especially on its hardest tests of frontier maths, long software projects, model development and visual questions (Tencent).

What changed from Hy4 preview

The preview, released on August 28 with downloadable weights, was a text-only model; WorkBuddy handed image and video jobs to other models (IT之家). Hy4 keeps the price and context and adds the rest:

Hy4 previewHy4
ReleasedAugust 28, 2026November 4, 2026
InputsTextText and images
Computer useNoCommand line and graphical interface
Context, max input, max output1,024K, 960K, 64K1,024K, 960K, 64K
TokenHub list price per 1M tokens6 / 18 yuan6 / 18 yuan
WeightsApache 2.0 on Hugging FacePlanned, no date
APEX-Agents v1.1, Tencent's run53.975.2
Terminal-Bench 4.0, Tencent's run26.344.4

Sources: TokenHub Hy call guide, TokenHub prices and the appendix on Tencent's release page. The preview's card lists 770 billion parameters with 49 billion active per token, a 1M-token window built on sparse attention with IndexCache, and an official FP8 build. Tencent's Hy4 page gives no parameter count; 科创板日报 and InfoQ report the Hy4 series at 770 billion. Tencent also reports that search-specific training cut its own hallucination measure from 15.76% to 5.45%, and errors from guessing from 23.1% to 5.6%.

Hy4 or Hy3

Hy3, released in July, is still on sale and is the cheaper Tencent model:

Hy3Hy4
Size295B total, 21B active770B in the Hy4 series, per the press
Context, max input, max output256K, 192K, 128K1,024K, 960K, 64K
InputsTextText and images
Computer useNoYes
TokenHub price per 1M tokens1 / 4 yuan, 0.25 cached6 / 18 yuan, 0.3 cached
WeightsApache 2.0 on Hugging FacePlanned

Sources: Hy3 card, TokenHub Hy call guide, TokenHub model list and TokenHub prices. Hy4 costs six times as much for input and four and a half times for output. In InfoQ's side-by-side runs Hy4 took 20 minutes 31 seconds on a business-plan deck against Hy3's 18 minutes 57 seconds and produced sharper strategy, and its posters showed more original ideas (InfoQ). For text-only jobs where speed and price matter more, Hy3 remains the sensible Tencent choice.

Where Hy4 ranks on independent leaderboards

The news since launch is Mercor's APEX-Agents board, which grades agents on 240 tasks written by bankers, consultants and lawyers, using expert rubrics and an AI judge. Hy4 at high effort completes 77.6% of tasks on the first try, plus or minus 4.5 points, third of all models listed (Mercor). The board has no open-weight filter and shows no date, so reports that Hy4 tops an open-weight APEX-Agents ranking go further than the board does: Hy4 does lead every model whose weights can be downloaded, but its own weights are not out.

Leaderboard (read November 6)Hy4Kimi K3.1 ProDeepSeek V4.1 Flash 1017GLM-5.3Claude Opus 5.5
APEX-Agents pass@1 (Mercor)77.6%65.5%69.8%56.6%73.5%
Vals Index v2.2 (Vals AI)59.35%62.52%61.26%59.73%70.18%
Intelligence Index (Artificial Analysis)3546.532.631.759.1
Terminal-Bench 4.0 (Artificial Analysis)37.4%67.2%33.8%41.9%59.6%

Sources: Mercor, Vals Index (updated November 5) and Artificial Analysis, each at the highest effort setting listed. Among the 19 models Vals labels open-weight, Hy4 ranks fourth, behind Kimi K3.1 Pro, DeepSeek V4.1 Flash 1017 and GLM 5.3; Vals labels Hy4's weights open although Tencent has not released them (Vals model page). Vals puts Hy4's cost at 1.72 USD per index task against 8.52 USD for GLM 5.3, for 0.38 points less. Artificial Analysis rates Hy4 below the median of comparable models, 35 against 37, and the most verbose: 160 million output tokens to run its index against a median of 72.6 million, 13.9 minutes per task and 103 seconds before the first answer token.

Tencent's own benchmarks

Tencent's appendix compares Hy4 with five models on 34 tests, 20 of them internal (Tencent). Where a rival has two scores, the first is a previously published figure and the second Tencent's own run; the table gives the first:

Test (Tencent's table)Hy4Kimi K3.1 ProDeepSeek V4.1 Flash 1017Claude Opus 5.5
APEX-Agents v1.1 pass@175.267.272.173.5
OSWorld 2.1, computer use78.979.668.481.8
DeepSWE v1.174.975.577.074.2
Terminal-Bench 4.044.465.742.366.4
ProgramBench, almost solved30.031.529.165.0
MMMU-Pro, images78.684.176.788.2
Goal2Repo (internal)23.221.020.374.0

The pattern: Hy4 leads this group on APEX-Agents and sits close to the others on software repair, trails Kimi K3.1 Pro on terminal work and image questions, and is far behind Claude Opus 5.5 on building a whole project from a spec. Mercor's own APEX-Agents score for Hy4, 77.6%, is higher than Tencent's 75.2, while the independent Terminal-Bench 4.0 runs, 37.4% at Artificial Analysis and 39.4% at Vals (OpenRouter), are lower than Tencent's 44.4.

What public tests found

Long tasks get finished, slowly. InfoQ worked with Hy4 inside WorkBuddy on launch day (InfoQ). A tower-defence web game went from a one-line request to a published page in about 55 minutes over three rounds; Hy4 wrote its own simulations to balance difficulty and checked the page in Chrome. InfoQ found later waves turned into idle clicking. A 12-page business-plan deck took 20 minutes 31 seconds against 18 minutes 57 seconds for Hy3, with deeper strategy but text too small to present. A 科创板日报 reporter's finance task, comparing Tencent, Alibaba and Baidu filings since 2020, ran 41 minutes 51 seconds and returned a 12-page PDF, 8 charts and a workbook with 183 formulas, close to an intern's first draft (科创板日报).

Images are read carefully, with slips. InfoQ's two-robot scheduling map came back with a collision-free plan and a proof that eight steps is the minimum. On an order form against a receipt, Hy4 matched every line and noted the inspection was still pending, but miscounted the mismatched items and claimed the record was handwritten without evidence.

Code costs time. AI/ML API, which sells Hy4, reports all 23 of its default runs passed, but a coding task took a median 4.4 minutes and about 18,000 reasoning tokens, against 25 seconds for DeepSeek V4.1 Flash (AI/ML API). Setting reasoning_effort to low kept all 8 coding tasks passing at 0.039 USD each; turning reasoning off failed 3 of 10. On ten questions about things that do not exist, Hy4 said it did not know on eight and ran out of tokens without answering on two.

What it costs

Tencent kept the preview's price. One million tokens:

RouteInputOutputCached input
TokenHub list price6 yuan18 yuan0.3 yuan
TokenHub billed now, half the list price until December 31, 20263 yuan9 yuan0.15 yuan
TokenHub batch, list price3 yuan9 yuan0.15 yuan
OpenRouter now, marked 50% off, no end date given0.417 USD1.25 USD0.021 USD
OpenRouter before its discount0.834 USD2.501 USD0.042 USD
AI/ML API0.882 USD2.646 USD0.044 USD
Nous Portal, November 60.17 USD0.50 USDNot listed

Sources: TokenHub prices, OpenRouter models API, AI/ML API, Nous Portal. The TokenHub discount ends at 23:59:59 Beijing time on December 31, and Singapore is billed at the same yuan rates as Guangzhou. TokenHub's price page shows no peak or off-peak rate for Hy4, as it does for DeepSeek models, so there is no night-time discount; OpenRouter lists a night rate, 10% lower between 00:00 and 08:00 Beijing time, only for Hy4 preview. Nous Portal, a reseller, does not say how long its price lasts. AI/ML API bills reasoning tokens as output, which matters for a model this verbose: Artificial Analysis spent 1.03 USD per index task at 0.834 and 2.501 USD per million tokens.

Tencent also sells Hy Token Plan subscriptions from 28 yuan a month for 560 points, which cover Hy4 and count its usage at a quarter of the usual points for a limited time (Token Plan). WorkBuddy and CodeBuddy give a free quota from November 4 to 15 (科创板日报): in WorkBuddy a small daily allowance until 23:59 on November 15, larger for subscribers, after which usage draws on points (快科技, 站长之家). WorkBuddy's night-time free window, 23:00 to 08:00, belonged to Hy4 preview in WorkBuddy and was set to run until October 31 (IT之家); Tencent's Hy4 announcements mention no night window. TokenHub's free-trial list includes Hy3 preview but not Hy4 (Token Plan). Tencent says a cheaper Hy4 flash is coming, without a date.

Weights, licence and running it yourself

Hy4's weights are promised, with a system card, and Tencent says it is working with the vLLM and SGLang projects and Inferact on support (Tencent). Until then you can download only Hy4 preview, on Hugging Face, ModelScope, GitCode and CNB. Its licence file is plain Apache 2.0 (LICENSE), which matters because older Tencent models came under the Tencent Hunyuan Community License, which did not apply in the EU, the UK or South Korea and required a separate licence above 100 million monthly users (Hunyuan-A13B licence). Tencent has not said whether Hy4 keeps Apache 2.0.

Running the preview takes a data-centre node. The official FP8 build is about 814 GB of files and the BF16 version about 1,560 GB (FP8 weights); vLLM's recipe serves it on 8 Nvidia B300 GPUs, 16 B200s or 8 AMD MI355X accelerators (vLLM Recipes).

Privacy, data and safety

Tencent Cloud's TokenHub FAQ says customer data is used only to provide the service and never for training (TokenHub FAQ). TokenHub sells Hy4 in its Guangzhou and Singapore regions (TokenHub prices). OpenRouter routes every Hy4 request to Tencent, which it lists as headquartered in China with a data centre in Singapore (OpenRouter providers), and says Hy4 is not available through its EU in-region routing. Buyers with data-residency rules should read Tencent Cloud's own terms before sending regulated data.

On safety, Tencent says it trained Hy4 only on defensive security data such as proof-of-concept and patch work, left out exploit development, reverse engineering and penetration testing, and had its Xuanwu and Keen security labs audit the result. On its own biology test Hy4 scored 30.4 out of 100, which Tencent reads as below the risk threshold (Tencent). Both checks are Tencent's own; the promised system card should give more detail.

Who should not use Hy4

Skip Hy4 if you need to self-host the final model now: only the preview's weights exist. Skip it for terminal-heavy coding agents, where Kimi K3.1 Pro and Claude Opus 5.5 lead by over 20 points on Terminal-Bench 4.0, and for interactive coding where a reply in seconds matters: a median coding task took 4.4 minutes. Teams that need EU-routed inference or a choice of hosts should wait, since OpenRouter has one provider, Tencent. And budget for the price after December 31, when TokenHub's 50% discount ends.

Alternatives

ModelInput per 1M tokensOutput per 1M tokensCached inputPrice page
Tencent Hy4 (OpenRouter, before its 50% discount)0.834 USD2.501 USD0.042 USDOpenRouter
Kimi K3.1 Pro2.00 USD10.00 USD0.20 USDMoonshot AI
DeepSeek V4.1 Flash, peak0.30 USD1.20 USD0.006 USDDeepSeek
GLM-5.31.40 USD4.40 USD0.26 USDZ.ai
LongCat-2.5, launch price0.30 USD1.20 USD0.006 USDLongCat
Mistral Large 40.70 USD2.10 USD0.07 USDMistral AI

Kimi K3.1 Pro is the better choice for coding and terminal agents and the strongest open model on the independent indexes, with weights on Hugging Face since November 2. It costs about 2.4 times Hy4's list price for input and 4 times for output.

DeepSeek V4.1 Flash is the better choice when speed and price matter more than the last points on agent tests. It reads images, its weights are MIT-licensed on Hugging Face, and off-peak it costs half its peak price.

GLM-5.3 scores about the same as Hy4 on the Vals Index and its weights are on Hugging Face, but it reads text only and cost five times as much per Vals task.

LongCat-2.5 from Meituan costs the same as DeepSeek V4.1 Flash at peak, under a launch discount, and also reads images and video, but its weights are not out either and its benchmarks are Meituan's own.

Mistral Large 4 is the choice for teams that want a European maker and downloadable weights, at a lower price than Hy4's list, though it scores lower on Mercor's APEX-Agents (51.6%), the Vals Index (56.53%) and Artificial Analysis's index (26.1).

If you only need Hy4 inside an app, Tencent's free Yuanbao assistant and WorkBuddy offer it without an API key; for the preview through many hosts, OpenRouter lists seven providers.

Sources

Sources

  1. Hy4 正式发布, Hy4 release page with benchmark appendix (Tencent Hunyuan, November 4, 2026) (hunyuan.tencent.com), accessed
  2. About Tencent: founded 1998, headquartered in Shenzhen, maker of Weixin/WeChat (tencent.com), accessed
  3. tencent/Hy4-preview model card: 770B total, 49B active, 1M context (Hugging Face) (huggingface.co), accessed
  4. Hy4 preview LICENSE file: Apache 2.0 (Hugging Face) (huggingface.co), accessed
  5. tencent/Hy4-preview-FP8 official FP8 weights (Hugging Face) (huggingface.co), accessed
  6. tencent/Hy3 model card: 295B total, 21B active, 256K context (Hugging Face) (huggingface.co), accessed
  7. Tencent Hunyuan Community License for Hunyuan-A13B, with its territory clause (Hugging Face) (huggingface.co), accessed
  8. TokenHub model list: Hy4 capabilities and limits (Tencent Cloud, updated November 5, 2026) (cloud.tencent.com), accessed
  9. TokenHub model prices and the Hy4 limited-time discount (Tencent Cloud, updated November 5, 2026) (cloud.tencent.com), accessed
  10. TokenHub Hy call guide: Hy4 protocols, reasoning and image input (Tencent Cloud, updated November 4, 2026) (cloud.tencent.com), accessed
  11. TokenHub product updates: Hy4 preview added August 28, Hy4 added November 4, 2026 (Tencent Cloud) (cloud.tencent.com), accessed
  12. TokenHub FAQ: does TokenHub train on my data (Tencent Cloud, updated September 3, 2026) (cloud.tencent.com), accessed
  13. Tencent Cloud Token Plan: Hy Token Plan prices and the Hy4 points offer (cloud.tencent.com), accessed
  14. 元宝 (Yuanbao) on the China App Store: version 2.92.0 adds Hy4 (Tencent, November 6, 2026) (apps.apple.com), accessed
  15. 腾讯混元 Hy4 正式版发布 (IT之家, November 4, 2026) (ithome.com), accessed
  16. 腾讯混元 Hy4、Hy4 flash 即将发布 (IT之家, October 29, 2026) (ithome.com), accessed
  17. WorkBuddy adjusts the Hy4 preview free period; preview is text only (IT之家, September 9, 2026) (ithome.com), accessed
  18. 腾讯史上最强开源模型来了:混元Hy4正式版发布 (科创板日报, November 4, 2026) (cls.cn), accessed
  19. 腾讯WorkBuddy首发接入混元Hy4大模型:限时免费体验 (快科技 on 新浪科技, November 4, 2026) (finance.sina.com.cn), accessed
  20. 限时免费!腾讯WorkBuddy首发接入混元Hy4大模型 (站长之家, November 4, 2026) (chinaz.com), accessed
  21. WorkBuddy extends Hy4 preview night-time free use to October 31 (IT之家, October 2, 2026) (ithome.com), accessed
  22. 混元 Hy4 实测:腾讯的产品基因,模型继承了几分? (InfoQ, November 4, 2026) (163.com), accessed
  23. APEX-Agents leaderboard (Mercor) (mercor.com), accessed
  24. Vals Index leaderboard, version 2.2 (Vals AI, updated November 5, 2026) (vals.ai), accessed
  25. Hy4 benchmarks, cost and capabilities, with rankings among open-weight models (Vals AI) (vals.ai), accessed
  26. Tencent Hy4 intelligence, performance and price analysis (Artificial Analysis) (artificialanalysis.ai), accessed
  27. Tencent Hy4 review: specs, pricing, benchmarks and test runs (AI/ML API, checked November 4, 2026) (aimlapi.com), accessed
  28. Nous Portal model prices, Tencent: Hy4 (Nous Research) (portal.nousresearch.com), accessed
  29. Tencent: Hy4 on OpenRouter, prices, provider and benchmarks (openrouter.ai), accessed
  30. OpenRouter models API: current Hy4 and Hy4 preview prices, including time-of-day rates (openrouter.ai), accessed
  31. OpenRouter provider list: Tencent, headquarters and data centre (openrouter.ai), accessed
  32. Tencent: Hy4 preview on OpenRouter, seven providers (openrouter.ai), accessed
  33. Hy4-preview serving recipe and hardware (vLLM Recipes) (recipes.vllm.ai), accessed
  34. Kimi API model inference pricing, Kimi K3.1 Pro (Moonshot AI) (platform.kimi.ai), accessed
  35. moonshotai/Kimi-K3.1-Pro model card (Hugging Face) (huggingface.co), accessed
  36. DeepSeek API models and pricing, deepseek-flash (DeepSeek) (api-docs.deepseek.com), accessed
  37. deepseek-ai/DeepSeek-V4.1-Flash-1017 model card, MIT licence (Hugging Face) (huggingface.co), accessed
  38. Z.ai pricing, GLM-5.3 rates (docs.z.ai), accessed
  39. zai-org/GLM-5.3 model card, text generation (Hugging Face) (huggingface.co), accessed
  40. LongCat-2.5 models and pricing (LongCat API Platform) (longcat.ai), accessed
  41. Mistral AI API pricing, Mistral Large 4 (docs.mistral.ai), accessed
  42. mistralai/Mistral-Large-4-1T-A52B weights (Hugging Face) (huggingface.co), accessed

Changelog

Changelog

  1. Page created from Tencent's own pages (Hy4 release page and benchmark appendix, Hy4 preview model card and licence, Tencent Cloud TokenHub documentation and prices), launch coverage by IT之家, 科创板日报 and InfoQ, the Mercor APEX-Agents, Vals Index and Artificial Analysis leaderboards, test runs by InfoQ and AI/ML API, and the price pages of five rival models.

Recommended from GenGrades

  • Tool review

    Runable

    Runable is a general AI agent that builds websites, apps, slides, documents, images and videos and now also runs ads and outreach, made by a 15-person startup in Bengaluru founded in 2025 by Umesh Kumar and Saksham Sarda. It turned on paid plans in March 2026 and raised a 21 million USD Series A on August 26, 2026.

  • Tool review

    Polsia

    Polsia is an AI service from Polsia, Inc. that says its agents plan, code and market a company with no human staff. Founder Ben Broca launched it in December 2025 and reports 10,000 paying customers; critics doubt his revenue figures, and his own numbers show nine in ten companies on it have yet to earn a dollar.

  • Tool review

    Browser Use

    Browser Use is an MIT-licensed Python library and a paid cloud service that let AI agents operate real web browsers, made by a Y Combinator startup founded by Magnus Müller and Gregor Žunič. Its GitHub repository had about 118,300 stars on October 17, 2026, up from about 50,000 at its 17 million USD seed round in March 2025.

  • Tool review

    WorkBuddy

    WorkBuddy is Tencent's AI agent for office work, built by the Tencent Cloud CodeBuddy team and launched in China on March 9, 2026. In seven months it added a global edition and phone apps, and Tencent announced a Guangdong government version. It plans tasks, works on your local files and hands back documents, spreadsheets and slides.

  • Tool review

    Raccoon Work

    Raccoon Work (商汤小浣熊) is SenseTime's AI assistant for office work: data analysis, documents and slides on web, desktop and phone. SenseTime merged its office and coding assistants into it on August 25, 2026, and its first-half results report cloud monthly users up nearly tenfold. No independent test of the new version is public yet.

  • Tool review

    Muse

    Muse is Meta's personal AI agent: it books, shops, sends email and fills in forms from its own cloud computer, asking before it pays or sends. Meta launched it in the US on September 8, 2026 and has since added Canada. Early reviews are mixed, and Amazon blocks it from buying on its store.