Hunyuan Is the Kimi Flash We Always Wanted

"Solving the Dilemma"

A while back, a secondary-market buddy hit me with a genuinely startling take:

It would be terrifying if Tencent actually figured out large models — bad news for every LLM startup.

Because Hunyuan was essentially built from scratch by Yaosheng Shunyun in less than a year. What this proves is that training large models has no real technical moat. If a big company commits resources, delegates authority to people who know what they're doing, and just grinds through the dirty work, they'll inevitably produce something usable.

This demonstrates that training large models is purely a resource game. And ByteDance, Alibaba, and Tencent obviously have more resources than the "Four Little Dragons." So Hunyuan succeeding is purely bullish for big-tech models.

Of course, this is a crude, sweeping generalization with plenty of counterexamples. The most obvious being big-tech organizational dysfunction — horse races, siloed departments, bosses who don't understand the technical details, hordes of algorithm bros who aren't even at Tian Keyu's level constantly competing for status and flooding leadership's brain with garbage data.

It's like that stupid question VCs love to ask founders: "What if ByteDance does exactly what you're doing?"

This used to be hard to answer, because VCs had already assumed ByteDance was invincible and would inevitably crush you.

Startup founders waited and waited for ByteDance to come for them, ready to finally curse out those stupid investors — only to check their WeChat Moments and discover those same stupid investors had already pivoted to even less falsifiable stuff like RSI and world models.

Now the question is easy to answer. Because ByteDance won't mobilize the entire company to destroy you. At most they'll assign a 3-2 level employee to try. And the resources available to that person may not exceed what a founder can muster, while their flexibility is far worse.

Melvin Chen proved exactly this.

LibTV made Dreamina miserable from top to bottom, yet Dreamina couldn't retaliate against Melvin. Because Melvin could wage price wars, adjust pricing at will, while big-tech employees had OKRs protecting gross margins and couldn't solve cross-departmental problems.

Volcano Engine sales and Seedance bros would sell to anyone — no incentive to give internal teams sweetheart deals. So you get this absurd situation where ByteDance's internal Seedance pricing might actually be higher than what Melvin publicly charges for memberships.

So unless Yiming Zhang personally intervenes and tells Dreamina to ignore profitability, Melvin's startup has no reason to die. And if your startup gets personally struck down by the Emperor himself — that's honorable defeat, glorious even!

I'm getting sidetracked. Today we're mainly talking about models.

Qwen is a classic counterexample. Everyone would say Alibaba is more "old-guard" than ByteDance, and there's no reason Auntie Wu understands large models better than Yiming. Yet Alibaba's base model Qwen is unambiguously ahead of ByteDance's Seed in performance — makes no sense.

Seed-Evolving has been pushing small updates, and newer versions have improved considerably. Especially since they barely distill, Seed's outputs feel genuinely human — no wonder I love chatting with Doubao so much.

But overall it's still second-tier among domestic models. Actually building products on Seed is pure masochism.

Let me add a caveat here. I know people keep saying Seed underperforms because it doesn't distill. But distillation is a means, not an end. If you're actually good, you should be good without distilling.

But on marketing, Qwen gets absolutely demolished by Kimi.

From K3 to K3.1 Pro, you can see Yutong Zhang constantly defining scenarios. K3's one-shot mini-game generation was a massive success, helping Yutong raise $10 billion.

Meanwhile, Qwen3.8-Max released around the same time was basically invisible externally — not even in the same league for buzz. But based on our rigorous evaluation, outside of frontend scenarios, K3 and Qwen3.8-Max are at the same capability level.

Qwen losing the marketing battle to Kimi makes perfect sense. Because it's actually Yutong versus some Alibaba P8, and that P8 really has it rough — every model release means taking Yutong's full-force assault head-on.

Yutong can chain together algorithms and operations, connecting post-training and marketing, defining user scenarios during post-training to hack global mindshare. So Yutong genuinely has something divine about him.

Divinity is the ability to chain people together.

An Alibaba P8 definitely can't command algorithm bros, so they're stuck reverse-engineering marketing scenarios from preview API versions. That they've managed this well is already impressive. I admire Yutong's creativity, but I respect the P8's stress tolerance even more. Bro's really been through it 🫡

In big-tech speak, this is insufficient pressure per unit area. Alibaba's total resources far exceed Kimi's, but distributed across every link and specific role, the resources any individual position can mobilize naturally fall short of Yutong's concentrated focus.

When will everyone realize that large model launches are no different from new energy vehicle launches — those launch days determine user mindshare.

There are so many models now, updating so fast, with entirely new batches every two months. Nobody will remember you.

Besides DeepSeek's transcendent brand value — whatever they release, even if Pro underperforms, everyone responds with support and encouragement — other model companies should really learn from automakers, who have rich experience rotating between ICU and KTV states, with exceptional stress tolerance.

Also, it's bizarre that nobody's organized a dedicated large model launch event yet.

Getting sidetracked again, back to Tencent Hunyuan.

Not much to say about the Hy4 model. First, it's a usable model. Based on my testing, Hy4 is about 80% of DeepSeek V4.1 Flash level — long-horizon tasks and relatively complex physics problems still struggle, easily spiraling into overcomplication, far behind Qwen and Kimi here.

But it's very fast, actually hitting 150+ tokens per second, plus it's practically free on Workbuddy. For most office tasks it's genuinely fast, good, and cheap — a qualified, usable Flash model.

Hy4 also aggressively optimized frontend capabilities. The MG animated videos it generates are quite polished; I couldn't tell them apart from K3's output. Shows Yaosheng really delivers — whatever scenarios are trending in the market, he'll train them in.

That said, objective assessment: Hy4 and DeepSeek V4.1 Flash both have a massive gap versus closed-source models — they're not the same thing.

Yaosheng's internal benchmarks quite pragmatically show this gap.

You don't even need to look at the scores. My experience with the Opus series is fast, precise, ruthless. From video editing to running benchmarks, as long as I describe my intent clearly, there's no unachievable task.

With these Flash models, you easily spiral into overcomplication, very mentally draining.

But Opus isn't available, so ChatGPT is still the convenient, usable option. The model-training folks are all grinding on ChatGPT too.

If you run these models side by side, you see a funny scene. Kimi, DeepSeek, and Hy all have a distinctly GPT-like frontend flavor — their outputs feel completely identical.

Considering completeness and cost-performance, we can say:

K3.1 Pro is like GPT-6 Pro, DeepSeek V4.1 Flash is like GPT-6, Hy4 is like GPT-6 Flash.

Hy4's actual moves demonstrate this too: Kimi's promotional video distilled from Claude; Hunyuan's promotional video distilled from Kimi. Kimi 3.1 Pro's three main scenarios are 3D, gaming, and finance; Hy4's main scenarios are also 3D, gaming, and finance.

Kimi as Claude substitute, Hunyuan 4 as Kimi substitute. Is this recursive self-improvement?

Back to the beginning: Yaosheng proved with actions that with resources and trust from leadership, you can indeed train a fully usable model from scratch in a year.

Put yourself in his shoes — this is incredibly difficult. Because big-tech work tests not just business competence but political competence.

Big tech hires a business leader for innovation, they fight fierce battles with old-timers, ultimately fail, leadership loses patience and shows them the door — sounds very familiar. The most famous previous case was probably Qi Lu at Baidu.

And Yaosheng, able to build large models from scratch at Tencent — a company famously good to employees, especially veteran employees — demonstrates maxed-out execution and operational talent, youth Zhang Juzheng level.

Qingliu that can actually execute crushes the turbid old guard!

Side note: Ming Dynasty 1566 is a great TV drama, and many people actually use it to study big-tech survival.

But another founder who's recently blown my mind is Tianxiang Sun. Brother Sun just released the new version of ERNIE Bot.

For those unfamiliar with ERNIE Bot, this is a magical model.

Previously when we ran benchmarks, a major frustration was domestic models being hard to differentiate — especially on coding tasks, that blob of Chinese models all clustered together, separated by mere decimal points, only roughly tierable.

ERNIE Bot was incredible because this model scored zero on most tasks. Input context too long? Direct error. Extremely differentiated, extremely convenient for publishing benchmarks.

Even more impressive, ERNIE Bot was China's first 2.4T parameter model. Imagine how much garbage data they trained into this thing.

Baidu acquiring Tianxiang Sun to retrain the model — few were optimistic from the start. When I learned ERNIE Bot was about to release a new model recently, I assumed Brother Sun was being sacrificed.

Because Baidu was about to hold a major conference, and Sun was likely to be trotted out for a big news moment — benchmark fireworks, then team explosion, reputation destroyed, no chance to startup afterwards.

Unexpectedly, Brother Sun showed flashes of Young Yaosheng energy. When he released the new model a few days ago, he directly compared against Hy3-preview-Base and DeepSeek-V4-Flash-Base — leaving Baidu PR zero opportunity for fireworks, pragmatic and grounded to the extreme.

I've already started an ERNIE Bot fan club group. Members are all praising Brother Sun's high level. If Brother Sun keeps at it for another year and gets ERNIE Bot close to latest DeepSeek Flash level, he can be canonized too.

Still needs observation though. Because Baidu's conference is in two days — there's still possibility of a big hype wave sacrificing Brother Sun. Don't turn him into Gao Hanwen.

Final summary.

The industry consensus I've encountered is that large models are grinding, dirty work. Technology diffuses fast, almost no secrets between model labs. So what most determines gaps between models is data, compute, and organizational capability.

Large model competition won't end in one or two years — it's destined to be a war of attrition. Data and compute as external resources will gradually follow market logic, matching to the cash on each model lab's books. This is why despite Seed being mediocre for now, everyone consistently bets on ByteDance.

So medium-to-long term, team stability — meaning the founder's ability to chain people together — far outweighs the importance of one or two bad releases, or falling behind by half a year.

Yaosheng proved at Tencent that you can crash-course a usable Flash model in one year. Next we'll see whether ByteDance Seed can sustain momentum after major reorganization, and whether Baidu's Brother Sun can survive a few more episodes. Without a company as entertaining as Baidu, I wouldn't know what to write about.

This leads to another question: For cloud providers, they inevitably face a choice between their own model or cloud revenue.

Google firmly chose cloud revenue; Gemini lagged for half a year. ByteDance firmly chose models, waiting for Yiming to chain the Seed bros together, for the Emperor's mighty hand to act. Alibaba in fact hasn't made a choice — externally, it appears Auntie Wu wants both Qwen and cloud revenue.

Tencent has a natural advantage: no need to consider cloud business. Tencent Cloud and Hunyuan aren't comparable things; the group unhesitatingly, unquestioningly supports Hunyuan.

Hunyuan has more resources than Kimi. Kimi, due to compute constraints, can only go all-in on SOTA models, effectively abandoning the consumer market. Kimi Flash going the fast-good-cheap route wouldn't match this company's divinity either.

So:

I see Hy4 is doing well, massive performance improvement, speed and cost gaps basically eliminated, reasoning capability also getting attention. If you added Yutong's management, Hy4 would be our ideal Kimi Flash!

⬇️

Subscribe to our Substack: funeralai.substack.com