The real answer to the Hy4 question is that Tencent is on this line of the model.Author of the book
Editor's note:
On November 4, the hybrid Hy4 was officially released, and the model's weight will be revealed in the upcoming open source, with a lightweight version of the Hy4 flash featuring a premium price tag. Tencent ranked it among the strongest open source models for the first time.
Over the past year, Tencent has been seen the most on AI, with agents and applications such as WorkBuddy; on the lower model, its volume is significantly weaker than applications.
In terms of performance on the list, Hy4 is already in first place in the open-source model; globally, the gap with the top closed-source model remains clear.
According to a recent set of data released by the National Bureau of Data, the number of daily Token calls in China has grown from 10 billion at the beginning of 2024, to over 500 trillion in July this year, more than 5,000 times in two years.
Polaris Park conducted tests around long-range Agents, multi-modal understanding and visual code, combined with official evaluation data and typical tasks, to determine what capabilities the Hy4 generation is complementing.
01
Long distance agent:
The threshold is "Run the whole thing".
In the past, the testing model, the industry's habit of looking at single-wheeled questions, single-way tasks. Going into the Agent stage, becoming a multi-wheeled, multi-step, long-running model, the model can't always remember where the goal is, which step it can go to, mid-way mistakes it can't pull back on itself, and finally form a logical closed-loop delivery.
Recently, with more focus on CPO racing, I've given Hy4 a task that's closer to the real buyer:
Let it act as a researcher covering the AI computing industry chain, around the CPO (Combined Optics) trajectory, to create an in-depth research report and recalculable valuation model for the investment decision-making committee, with data up to the most recent completed trading day.
The year 2026 is known as the year of mass production of CPO, with CPO exchangers such as Spectrum-X, Bailly and Bottling continuing to enter customers, COUPE production, Google NPO order placement, and industrial events; at the same time, A-share optical communications ("Easy Midnight" ("Easy Midnight") ("Mid-Times") ("Mid-Times") ("Mid-Times") ("Mid-Times") ("Mid-Times") ("Mid-Times") ("Mid-Times") ("Mid-Times") ("Mid-Times"), "Mid-Times" ("Mid-Times"), "Times") and the share prices and valuations of optical chip and optical equipment companies have risen sharply.
The task required coverage of almost the entire workflow of a researcher. It first dealt with CPO's industrial chain mapping of upstream light chips, lasers, filters, optical devices, midstream light engines, optical modules, downstream switches and cloud manufacturers, along with a visualized industrial chain map; while explaining the technological differences between CPO and traditional plug-in light modules, NPO, OUI, alternative rhythms, and key catalytic nodes, citing the official roadmaps and timing points of the U.K.
Video source: Park of the Poles
Data verification is the most revealing aspect of the topic. The model needs to obtain the most recent three-year financial statements of key companies, the latest trends with market capitalization, business structure, gross interest rates and institutional expectations, as well as overseas mapping data such as Intel, Bottling, Coherent, Lumentum; each digit must identify sources, dates, currencies and units, distinguish between actual values, forecasts and assumptions, lack of data clarify, prohibit the creation of data, or the identity of analysts. It also specifically requires that it verify whether our companies have actually entered the overseas giant CPO supply chain, strictly distinguish between production volumes, sample testing and demonstration cooperation in three states.
On this basis, it will select four companies representing A shares that cover optical modules, optical devices, and optical chips, and measure the growth expectations of the current valuation implicit, combining consistent expectations to determine where the expectations differ, giving an optimistic, benchmark, and pessimistic three-dimensional scenario; then select a company that looks to build a recalculable future three-year profit model, key assumptions for centralized management, verification of collusion, and prohibition of unexplained overcrowding.
From a practical point of view, Hy4 follows the typical path of dealing with long tasks: first dismantling tasks, listing execution plans, phased progression, and intermediate stages of self-checking. Industrial chain diagrams, technical routes, comparison tables and three scenarios are clearly marked in sequence, data sources, assumptions of quality and uncertainty, correlation within the model, and the final delivery is to report PDF value-added model xls.
The official financial crowd measurement data can be used as a reference, with Tencent building thousands of workspaces and tens of thousands of high-quality tasks based on real financial scenarios, covering processes such as fund research, securities analysis, quantification strategies, asset allocation and tax accounting, and jointly with Tencent Financial Technology FiT organizing 213 expert crowd measurements, of which 153 were from outside leading financial institutions, seven with more than five years of experience. In 579 real financial missions, Hy4 received 3.04 points, ranking first in the open-source model, higher than DeepSeek-V4.1-Flash 2.97's KLMimi K3.1 Pro and G-5.3's 2.68; in the internal ratings of the financial intelligence company Tencent Technology IRAB, 4 was upgraded by 12.03 points from preview versions.
A further improvement was in fact accuracy. For time-efficiency detection, multicore information integration and fact consistency, mixed with special search optimization, Hy4's hallucination rate dropped from 12.95% in the preview version to 7.2%, and the error rate caused by speculation dropped sharply from 23.1% to 5.6%.
The gap in long-range capability is often hidden in the last kilometre. Ninety percent of the previous steps, most of today's first-tier models can do, can actually generate a file, can get the numbers right, can be opened, and that's the watershed. Hy4 in this circle has the shape of a "primary research assistant".
Of course, this type of task remains largely unacceptable. Models will mark the quality of the data, assume rationality and final conclusions, but the decision-making responsibility of the selection committee is still not something that any model can take on for the researchers.
02
Visual understanding and visual code:
The first blend of "eyes"
One of the most notable changes of the Hy4 generation was the introduction of visual understanding for the first time. It was preceded by the purely linguistic model route, which began to understand images, graphs, snapshots, and video, and convert visual information into output. According to official reports, it was able to complete these three tasks: fine-grain image processing, complex graph analysis, and mathematical visual reasoning.
We focused our testing on coding tasks that rely on visual restoration, which is also where multi-mode capabilities are most easily seen. The first task was an interactive 3D mechanical clock movement to demolish the web.
Requires the use of Three.js to display a programmed modeling mechanical clock that contains at least a clock, slide, mirror, dial, dial, clockwise clockwise clockwise clockwise clockwise clockwise clockwise clockwise clockwise clockwise clockwise clockwise clockwise clockwise clockwise clockwise clockwise clockwise clockwise clockwise clockwise clockwise clockwise clockwise clockwise clockwise clockwise clockwise clockwise clockwise clockwise clockwise clockwise clockwise clockwise clockwise clockwise clockwise clockwise clockwise clockwise clockwise clockwise clockwise clockwise clockwise clockwise clockwise clockwise clockwise clockwise clockwise clockwise clockwise clockwise clockwise clockwise clockwise clockwise clockwise clockwise clockwise clockwise clockwise clockwise clockwise clockwise clockwise clockwise clockwise clockwise clockwise clockwise clockwise clockwise clockwise clockwise clockwise clockwise clockwise clockwise clockwise clockwise clockwise clockwise clockwise clockwise clockwise clockwise clockwise clockwise clockwise clock clockwise clock clockwise clockwise clockwise clock clock clockwise clockwise clockwise clock clock clockwise clock clock clockwise clock clockwise clockwise clock clock clock clock clock clock clock clock clockwise clock clock clock clock clock clock clock clock clock clock clock clock clock clock clock clock clock clock clock clock clock clock clock clock clock clock clock clock clock clock clock clock clock clock clock clock clock clock clock clock clock clock clock clock clock clock clock clock clock clock clock clock clock clock clock clock clock clock clock clock clock clock clock c
Video source: Park of the Poles
To create such a web page, the model must first use a programmed geometry to put dozens of parts together one by one, attach each piece to a level and coordinate, move in order when disassembled, ensure the position of each other, and keep the gears and swingers running continuously. According to the official evaluation and task requirements, it can rotate, scale, dismantle, and reconstruct the entire set of interfaces.
The second task was to get it to develop an online version of the game called NAIAD, similar to the gameplay in the video game.
The film was released in theaters and on home video in the United States.
Production process of the video source: Polaris Park
As you can see, in this round of testing, Hy4 is actually a closed loop sequence of video proofs, mechanism abstraction and rendering re-reading: first use ffmpeg to tap the screen for 81 seconds, puzzles, local magnification, and repeatedly identify "white manfish moves, collects light points, touches flowers, flows through a circle, flows through a circle, chapter cards fade in, deaths countless times", and then use single file Canvas 2D to rewrite.
It's actually the last half of the process, with the headless browser running correctly, scanning, then reading the screenshot back and forth, continuously pulling out the wide-screen flower is cut, the width of the coordinates is changed, the light points are unsatisfied, the chapter door closes four problems, it runs again, cycle four or five rounds until the chapter progresses and the buttons are all passed. This is a "multiple reading → re-doing → rendering re-reading verification cycle: visual understanding" is only responsible for direction, making subjective judgments into verifiable self-reality, it's back checking.
In addition, computer operation is the ability to support Hy4 for the first time. It can see the screen like a human, and then use the mouse keyboard to operate the browser, desktop applications and operating systems, command line and graphical interface.
