Tencent Hy4
hy.tencent.ai293 points by Philpax 3 days ago
293 points by Philpax 3 days ago
> Looking ahead > Although Hy4 is one of the strongest open-weight models available today, a notable gap remains between it and top proprietary models, particularly on challenging internal benchmarks for research-level mathematics, long-horizon coding, model self-improvement, and advanced visual question answering. We will continue to steadily scale pre-training and post-training to narrow this gap.
I like that they acknowledge that. For working with ai as a tool open models are already great. For millennium problems and doomermaxxing they are not there yet.
What is that in English ?
Writing incoherently wastes your time, since you are generating an idiolect that nobody bothers to actually engage with.
> doomermaxxing
Is this the kind of quality verbiage I’m missing out on, not using Discord?
I misread it as "doommaxxing" and could only guess it somehow involved Doom (1993).
Yeah, never been happier sitting on Teamspeak and not getting hit with Zoomer-speak.
It is better to accept that language changes than becoming the old man yelling at clouds.
I don't think I've heard someone call something "radical" as a compliment since the aughts. "Language changes" is abused as a shallow defense of current-year-ese much more often than it's used as a useful descriptor of actual shifts.
Hey, I’m just friggin around - I wish I had the energy to sandbox and keep up with the flash in the pan that is Discord’s walled garden…
but…why did this hit you so hard?
The game development focus is interesting:
> For game development, we partnered with professional developers at Tencent Interactive Entertainment Group (IEG) to curate high-quality training data and build evaluations. We focused on feature implementation, gameplay logic, state tracking, and asset pipelines. We also extensively adapted Hy4 for major engines such as Unreal Engine, Unity, and Godot, and expanded its capabilities to multiplayer game development. Compared with Hy4 preview, Hy4 improves by 19.1% on GameDev and 9.8% on GameLogic.
Here's more about that game development studio - https://en.wikipedia.org/wiki/Tencent_Games - they've been around since 2004.
Tencent is one of the largest game development asset holders in the world. They're the Berkshire Hathaway of video games. They own, World of Warcraft (minority shareholder), League of Legends, Valorant, Warframe, Clash of Clans, Player Unknown Battlegrounds, FIFA Online, Pokemon Unite, Fortnite (minority shareholder) annd more. They also own 40% of Epic Games.
> They're the Berkshire Hathaway of video games
They are the Berkshire Hathaway of everything, to be honest. They invest in and acquire minority stakes in hundreds of companies.
Wikipedia has them at ~a fifth to ~a half depending on metric:
https://en.wikipedia.org/wiki/Berkshire_Hathaway https://en.wikipedia.org/wiki/Tencent
Do they have a stake in Fortnite somehow specifically/separate from their stake in Epic Games?
Instead of listing the publisher/studio (Epic, Riot, etc) he's just listing their games directly.
> They also own 40% of Epic Games.
I wonder if that has anything to do with the latest vibecoded slop releases and shilling from Tim Sweeney.
And they probably arent going to block decomp/cloning.
edit: apparently im wrong on that
Landed on the cost/accuracy frontier in our data analytics benchmark: https://plotly.com/blog/tencent-hy4-plotly-data-analytics-be..., right in between GPT-6.1 Sol and GPT-6.1 Terra.
Verrrry thorough in its work and therefore verrrrry slow. It spent a long time with each question, running about 26 queries for each question, and in some questions it ran 64 queries. One of the queries that it wrote also took about an hour to run, which affected its overall exam time.
Last year the open weight models were the cheap & good alternatives. Then in September/October that cost story got rewritten by the GPT-6 series and Claude's 5 series.
Hy4 brings open models back on the cost/accuracy frontier, but its behavior is too thorough + slow.
Another thing we're seeing is that post training open models is starting to see results. Particle-1 from last week is also on our frontier, which is post-trained on DeepSeek V4.1 and Fireworks Ember-1 from a few weeks ago was post-trained on Kimi K3.
Take these numbers with a grain of salt: this exam is just for short horizon, low-context data analysis tasks, not full-fledged software projects.
Slightly worse than Kimi 3.1M Pro (which just released open weights today) but much cheaper even excluding the current 50% off promotion, and Tencent promises to make it even cheaper once it goes open-weight. Interesting price economics.
Kimi 3.1 Pro is 2.8T vs this at 770B. I find that the jump in parameter count gives less tangible benefits like better "taste" and long context coherency, but it depends on the problem if that matters obviously.
increasingly feeling like the entire US economy hangs on whether or not Ant/OAI can stay ahead of the Chinese models.
If the gap closes, we do zero austerity, and interest rates are high, I cannot see a way out of a major economic crisis, even if we do start to see dramatic productivity wins.
More on-topic (and maybe more positive?), but this model feels a bit more expensive than I would have expected!
It's not the Chinese models' fault those two companies' business models are doomed. Quantization and model architecture improvements have gotten a few orders of magnitude win, and hardware has gotten another (and I'll bet another 10-100x is coming soon).
OpenAI/Anthropic remind me of Silcon Graphics. Sure, lots of people bought those workstations before MP3 / MPEG video were a thing, and before Voodoo came out with the first cheap 3D accelerators. All the tech trends (software and hardware) back in the 1990's were against them, but their valuation stayed high for a bit longer than rational.
Unless there's some second step function change in AI, LLMs comparable to today's state of the art are going to be running on $1-2K machines in a few years, and they'll be doing it fast and with low wattage. There are already a dozen companies that know how to build that, including in the US, Europe and China.
Unlike SGI: Hopefully, given all the political corruption and fraud in this space, we'll see jail time for the guilty parties.
SGI did try to pivot their business towards less expensive workstations, but they didn't manage to create a compelling product. The pivot might have been feasible because it would have stayed a hardware business.
But AI labs are datacenter businesses now with the AI part being the driver. Pivoting to selling AI weights for on-premises hardware is a much bigger change. Given the money tied up in data centers, this will be very tough times.
> LLMs comparable to today's state of the art are going to be running on $1-2K machines in a few years
This is a huge claim. What makes you think this will be the case?
Qwen 3.8 27B, not even the superior Flash Next 125B, running on a single 3090 in my workstation handily whoops ChatGPT 3.5 from just a couple years ago.
Seems like he's just drawing that line into the future. Which seems sensible enough.
I'm drawing the line into the past. Back when Biden was president, a computer that currently runs Flash Next 125B well cost $1400.
But yeah, I assume that when the economy gets back to normal-ish, in 2024-equivalent dollars, we'll be at 500B-1T for < $2000. That'll be 3-4 Moore's Law doublings since computer prices were sane. 125B * 2 ^ 4 = 2T.
> I cannot see a way out of a major economic crisis, even if we do start to see dramatic productivity wins.
The only thing worse for the economy than failure would be success.
> whether or not Ant/OAI can stay ahead of the Chinese models.
Ant/OAI/Google are way ahead when it comes to the smartest, largest-scale and most compute-intensive models. By Jevons' paradox, if you make inference for leaner models cheaper (as the Chinese models do, being broadly comparable to GPT Terra or Claude Sonnet; barely approaching Opus/Sol at their very best) that only increases demand for the complementary service of very smart inference.
Hard disagree. I own my gb200 cluster and have unlimited supply of open weight tokens. Whatever is missing from the open weights, they can make it up from speed and multi agent swarm.
Yeah: quantity is a quality all of its own. DS v4.1 Flash running at ~300tk/s makes its verbosity less of a downside, and the efficiencies the Chinese open model space has aimed for makes high concurrency easy for swarms/sub agent sets.
On the other side, I am still shocked at how good Qwen 3.8 Flash Next is on my GB10 Spark-like, even if it is only 45 tk/s decode!
Sure but routing simpler things to the open models and harder things to opus/astra gave us already a massive boost; rarely we run out of tokens on the subscription models and we recently cancelled a few seats on claude code / codex as we simply were underutilising. And also, we have users of our products ONLY using the cheap models as they benefit enough vs paying more. That was very different only months ago. I can see this gap just growing where the % of higher intelligent model use will get smaller and smaller.
> Sure but routing simpler things to the open models and harder things to opus/astra gave us already a massive boost
You have to look at the effect in the aggregate, not just what any single actor is doing. This is the core of Jevons' paradox: it's based on surfacing additional demand for the same product (here, expensive model inference) that wasn't previously viable.
> By Jevons' paradox
Jevons Paradox is not a rule its just a sometimes observed phenomenon
It says nothing one way or the other whether that would happen or not
There was always going to be an economic "crisis", just a matter of time. I see it more as a forced positive economic adjustment (bleak for those enamored with the current system).
Shut down stock markets, liquidate the failing companies, use that money to buy out stakes of pension and index funds, leave the debt with the company and let banks write it off. They're supposed to have insurance and risk controls anyway.
If any important bank fails because of this, nationalize it and let their customers exit using money from the federal reserve. Then try and recover as much of that as possible using the banks investment portfolio. Costs a lot less than bailouts.
> liquidate the failing companies, use that money
What money are you expecting to get int hat scenario? If I had to guess it wouldn't be a fraction of the cost.
> buy out stakes of pension and index funds
Index funds are part of the reason the top companies stock gets so bloated compared to those just outsid of their frame.
>If any important bank fails because of this, nationalize it and let their customers exit using money from the federal reserve. Then try and recover as much of that as possible using the banks investment portfolio. Costs a lot less than bailouts.
I agree. It worked here in Belgium tho they did spin off toxic elements of the bank in a not so great way prior to doing this.(some regular people still got shafted)
If you shut down the stock market the financial system craters. As in stone ages no more than economy.
Modern financial system is less „finite amount of money you can move around in a logical way“ and more confidence based definitely totally not a ponzi scheme
It's arguable that there is a threshold of quality that matter and US and China are both there already. I guess we'll see. Many of us on here are software engineers and I think we've seen that while Opus 4.7 (I think) was some sort of outlier/miracle, today there are several options to choose from. Depending on the task, the harness, etc Astra/Fable are only better when the work demands their level of capability - which much of software engineering does not.
> increasingly feeling like the entire US economy hangs on whether or not Ant/OAI can stay ahead of the Chinese models.
I’ve said it for a while now both here and to friends - by Christmas next year, US AI labs will be trying to catch up with Chinese open models.
… and I think once Qwen’s latest open model beats every US labs, it’s over.
I think whether or not there is an economic impact from AI is really independent of Anthropic and OpenAI.
If there is an economic boost, it will come from broad productivity improvements and deflation of costs for services. If OpenAI or Anthropic can capitalize on that or not is a separate question.
i dont see the connection from 2 model providers relative position to the entire economy? especially in the face of dramatic productivity wins.
Look back to the "too big to fail" banks in '08.
From biggest to smallest, the top three:
AIG controlled $1.05T and market cap crashed from $180B to $3.8B
Wachovia controlled $782B and market cap crashed from $110B to $1.8B
Lehman Brothers controlled $639B and market cap crashed from $60B to $200M.
Massive companies accounting for lots of money in people's retirment accounts. The companies made mistakes leading to what would have resulted or did result in defaulting on their debts. Suddenly billions of dollars vanished from retirement accounts.
We could very easily see a similar cascade of corrections if we get a few more "Alphabet missed earnings" news calls.
> the entire US economy hangs on whether or not Ant/OAI...
... What on earth. Please go outside for awhile.
Not cost competitive. The performance difference is not commensurate with the price.
Model | Cached Input | Input (Miss) | Output
---------------|--------------|--------------|----------
Tencent Hy4 | $0.042 | $0.83 | $2.50
DeepSeek (peak)| $0.006 | $0.30 | $1.20
DeepSeek (off) | $0.003 | $0.15 | $0.60We've moved past per token cost a long time ago. Smaller models use more tokens. What's interesting is the cost per task, on that front deepseek isn't particularly good.
The AA benchmarks seem to be mostly "complete this small intelligence-proving task from scratch". That consistently puts models with high intelligence to output tokens on the frontier, especially the OpenAI models, on the frontier. But a lot of software engineering work is many-turn work on established code bases. There's a lot of reading and a relatively large context unless it's carefully controlled. Writing only a few output tokens doesn't really cut costs because most of the spend is read tokens that get recycled through dozens of turns. That's basically the ideal case for DeepSeek because of how cheap cached tokens are.
It highly depends on the task. So far, on the tasks I personally benchmark, DS v4.1 Flash @ deepinfra.com is always cheaper because their cached input price for it is so very low and their cache hit rate is so very high.
do you generally favor deepinfra or you are mentioning them because it's 50% off on openrouter?
I really want to see some comparison on deepseek 4.1 flash 1017 providers, like, what's more trustworthy to not quantize without disclosing it, etc
also I can't find the official DeepSeek provider on openrouter https://openrouter.ai/deepseek/deepseek-v4.1-flash-1017
That analysis is unfortunately running DeepSeek on Max. DeepSeek recommends against that since it guarantees higher cost and rarely improves the outcome- it should be ran on "high".
They have "high" on there now too, just search for it.
Found it on that page, but they don't have it on their coding page.
Their coding page shows DeepSeek Max as Pareto optimal: https://artificialanalysis.ai/agents/coding-agents Costing 1/2 of Terra Xhigh. It would also be great to see Sol and Opus not on max there.
Not a useful benchmarking site for coding unless you only want to run your models on Max.
This benchmark is total bs and does not represent real world usage. None of the models that “beat” Deepseek in the benchmark are any good.
Wild that this is the first question I ask at this point. After that it’s a question of how it stacks intelligence wise vs glm or ds.
OAI / Ant are in for a reckoning.
How does it do with tasks anthropic doesn’t like? This is my main use case for open models.
>We filtered out offensive content such as exploit development, reverse engineering, and penetration testing to avoid amplifying potential offensive capabilities.
So, yeah. Back to GLM 5.3.
It's almost surprising that there's no big AI org that's doing the Star Wars "MORE!" meme in regards to increasing offensive capabilities. I guess everyone wants to play the US safety concern theatre. Or bad actors are just doing their best to jailbreak SOTA and dance around classifiers or throw 5x more non-SOTA tokens at issues.
I hope our Chinese brothers start naming their models in a more entertaining way, for example Deepseek 4.2 could Gwuiluo Surprise.
大哥哥 Model. Very caring. Sees everything.
大哥哥 is what a kid would call a much older kid for example.
The character's name is 老大哥. That is if we are talking about 1984.
The price numbers look weirdly specific, I wonder where they came from.
Personally, the thing I'm most excited about is the FrontierCode 2.0 score, which is a benchmark I trust a bit more - 31 seems to be roughly on par with 6.1 Sol and Opus 5.5 medium
Refreshing to see normal looking bar charts for the benchmarks for once. Yes it shows they are in the middle of the pack but at least we can understand.
Similar to Kimi K3.1 Pro, they nuked their cyber security skills too.
Similar to Kimi, they didn't "nuke" anything. They opted not to train the model on cyber-offense tasks which is a very reasonable choice when compared to what OpenAI and Anthropic do (train on cyber offense and then do a surprised pikachu face when their models go rouge and engage in cyber offense), and gives the model more scope for training on benign tasks.
Videos 1 and 4 on vision and computer use are wild. They depict the model finding a 2m feature from a large satellite image, then determining where a photo was taken. Is it just me or do the Chinese model releases have less concern about looking like a military weapon? I don't think it's a bad thing necessarily, it seems honest. Also the music they choose for it is hilarious.
China hit maximum tech dystopia a long time ago. Now they can enjoy tech because they know its already maximally invasive and there is nothing they can do.
Last time I went to China, a friend left her phone in a taxi (I guess a "didi") near a train station. She ran to the train station's police station, pestered them into looking at the recordings, got them to tell her the license plate of the car, and figure out how to call the driver. Her phone was back 20 minutes later.
If you're going to be surveilled 24/7, at least you gotta make it work for you.
> If you're going to be surveilled 24/7, at least you gotta make it work for you.
I don't get it.
There was CCTV at the station, your friend arrived in a taxi. And we're supposed to be amazed you can call up the taxi company and quote a numberplate off CCTV ?
Can be done anywhere in the world.
Would have been far more impressive story if she had arrived in a private car, dropped off by a work colleague or something.
> And we're supposed to be amazed you can call up the taxi company and quote a numberplate off CCTV
No. It was just a thing that happened to me that I thought was funny and interesting, because running to the train police station and asking them to track the didi was not something I'd ever have thought to do.
I'm sorry my little anecdote didn't satisfy you.
> I'm sorry my little anecdote didn't satisfy you.
It wasn't that. Its that you made a reference to the Chinese surveillance state in the same post. Your example didn't fit the evil-bogeyman narrative you were aiming for.
Sorry if I was unclear there.
> Your example didn't fit the evil-bogeyman narrative you were aiming for
What in the world are you talking about. If anything, I wasn't sure if recollecting that anecdote was minimizing the harm of being surveilled constantly. The last thing I expected was being accused of an "evil-bogeyman narrative."
> Sorry if I was unclear there.
Yes, thanks for clarifying. Now I understand that this has nothing to do with me or what I wrote and is 100% projection on your part.
> What in the world are you talking about.
You used the words "If you're going to be surveilled 24/7, at least you gotta make it work for you."
The implication being quite clear you were inferring it was only possible in China's surveillance state, when it was not.
How about we cordially agree to disagree and move on ?
> The implication being quite clear you were inferring it was only possible in China's surveillance state, when it was not.
Jesus fucking Christ.
Last time I did that, I called my phone from my friend's phone, asked the driver to come back, and paid him for the trip.
> figure out how to call the driver
Which country are you in where you cant take "call the driver" for granted?
Going from license plate to phone number would be one hell of an endeavor in most US states.
I expect the US police to have this capability; < 1% expectation that they would share it with a random citizen who walked in.
I don't quite understand the question. If she had had her phone, she could have called the driver using the information from the didi app, but she had no phone, so she didn't know who to call.
They can enjoy tech because people see tech improving every aspect of their lives. Nowhere in the the world have people ever seen the rate of technological progress that people in China have. Imagine going from having practically no electronics, to everyone having a cellphone, self driving cars, high speed rail across the whole country, drone delivery, and so on. Naturally, people who've seen technology dramatically improve every aspect of their lives would be optimistic that further technological progress will bring even better life.
Meanwhile, in the west, the standard of living has been eroding for the majority of the population. Technological progress is associated with layoffs, loss of agency at the workplace, and diminishing life prospects. So, it's no surprise that there is growing opposition to AI.
I can imagine it because the same thing has happened in my life time. The difference for me is that the tech isnt used to suffocate any of my human rights.
Tech in China isn't used to suffocate anybody's human rights either. In fact, more people in China see their country as being democratic than in America according to the latest Democracy Perception Index from the Alliance of Democracies Foundation.
https://allianceofdemocracies.org/democracy-perception-index
What makes this report interesting is that the organization publishing it was started by an ex-NATO chief and claims to unite democracies, meaning Western liberal style democracies. So it can’t be accused of having some sort of anti-Western bias. Their methodology is also straightforward, where they simply ask people if they feel their government works in their interest and whether they can hold it accountable. That’s the whole ballgame, because if people don’t believe the system is democratic then it inherently lacks legitimacy no matter what the constitution says. While people will surely attack this methodology, the sheer fact that a NATO-connected group published such a report gives its claims a certain credence given how dire these results are for the Western model of democracy. France now ranks alongside Kazakhstan, Yemen, and Zimbabwe in its own citizens’ eyes while the U.S. is split down the middle with about half of Americans saying they don’t think they live in a democracy anymore. China, on the other hand, sits up there with the Scandinavians and Switzerland, and Chinese people overwhelmingly say the system works for them. Notably, we can also see that perceptions of transparency, rule of law, and free speech are on the rise across the board in China.
And before you start trying to claim that people in China are simply afraid to speak their mind, the same report shows Russia at -21 points and Belarus at -9, so that whole excuse falls apart. Turns out that people living in states the West labels as being authoritarian do, in fact, complain when they’re unhappy.
Various other western publications have reported the same thing for many years.
https://www.tbsnews.net/world/china-more-democratic-america-...
https://www.newsweek.com/most-china-call-their-nation-democr...
https://www.csmonitor.com/World/Asia-Pacific/2021/0218/Vilif...
Turns out the US is the place where human rights are actually being suffocated.
> Tech in China isn't used to suffocate anybody's human rights either
Explain the mass uploading of secret screen recordings in more than ten phone brands sold in China in a component called SystemHelper, disclosed in the past few days, including but not limited to Xiaomi, Samsung, OPPO.
Ah yes, because American companies absolutely do not vacuum up all your data. Also, big news to me that Samsung is Chinese.
Whataboutism, so predictable.
Also, read again what I said, 'sold in China'.
Does it hurt so much when your fantasy turns out to be fake?
Oh, and you haven't explained the uploading. Maybe your last visit to onlyfans is now in a server, you know.
Screeching whataboutism is what low intellect individuals do to justify applying a double standard. What Chinese companies do in China is basically the same thing that Americans do in the US and everywhere else. Yet, people claim America is a free country. So, apply the same logic to China?
If you need to be explained why companies like to harvest customer data, then what else is there to say about you.