The high-priced Kimi K3, the cash-strapped Moonshot AI

The Kimi K3 was launched with massive fanfare. 

With 2.8 trillion parameters, a million-token context window, and native multimodality, Moonshot AI directly placed it in a ranking alongside Claude Fable 5, GPT-5.6 Sol, and Claude Opus 4.8 for comparison.

However, today's users are no longer swayed by ranking gimmicks. Right after K3 went live, community feedback quickly split into opposing camps.

Some users, after testing long codebase analysis and long-document reasoning, exclaimed that "the domestic flagship model has finally caught up," noting that the Agent cluster can work continuously for four hours without dropping tasks;

Others shared real-world testing experiences complaining that the model is overly verbose and the generated infographics lack aesthetic appeal. Some developers explicitly stated that their programming tasks failed to run properly on K3, and switching back to Claude Opus delivered far better results.

There are also widely consistent observations: the final outputs are occasionally impressive, but the model is two to three times slower than competitors, consumes tokens extremely quickly, and users burn through their subscription quotas in no time.

Many users commented that "it reeks of over-the-top marketing, and the actual user experience cannot live up to its claimed top-3 global status."

This is roughly the most accurate portrayal of K3's current state: it performs outstandingly on benchmark rankings, but its real-world work capabilities still need time to be fully validated.

The Awkward Position of Kimi K3

Setting aside promotional rhetoric, if K3 truly lives up to the official claims of its capabilities, the most comparable domestic model to it right now is Zhipu AI's GLM-5.2.

These are the two domestic open models that are most frequently put head-to-head against cutting-edge U.S. models at present.

GLM-5.2 was released one month earlier, and has already gone through a full round of validation via API calls, Coding Plan subscriptions, and developer toolchain testing. It has a far longer track record of real market feedback than the newly launched K3.

Many users refer to GLM-5.2 as a "cost-effective alternative to Claude for coding," and its Coding Plan offering is often in high demand with no available slots for many users.

More critically, there is a massive gap in pricing.

GLM-5.2 charges $1.4 per million input tokens and $4.4 per million output tokens; K3, by contrast, is priced at $3 per million input tokens and $15 per million output tokens, making K3's output price more than three times that of GLM-5.2.

This puts K3 in a very awkward position.

For the vast majority of enterprises and developers who need to strictly control costs, without a clear, measurable gap in performance, the price difference becomes the most practical deciding factor.

When it comes to affordability, it cannot compete with Zhipu AI's GLM series or DeepSeek; when it comes to raw capabilities and brand recognition, it is still no match for Claude or GPT for now.

K3 must prove that its 2-3x higher price tag compared to GLM-5.2 translates to a significantly higher task success rate.

Otherwise, why wouldn't users just stick with GLM-5.2, or even pay the premium to use Claude Opus 4.8 or Claude Fable 5? Why would they take on the added cost of migrating their workflows to K3?

To be fair, Kimi does have its own unique advantages.

It has a well-established consumer-facing brand, years of accumulated expertise in long-context model development, a full suite of products including search, Kimi Work, dedicated coding tools, Agent systems, and multi-agent frameworks, and a deeper understanding of everyday end-users than many pure-play model developers.

But all these advantages ultimately need to answer one core question:

What specific tasks are too complex for GLM to handle, too demanding for DeepSeek to execute well, and too expensive to run on Claude, where K3 is the only optimal choice?

Kimi's Capital Market Narrative

The launch of K3 at this juncture is not just a regular model upgrade—it is a carefully timed move to present a fresh growth story to the capital market.

Moonshot AI has been accelerating its fundraising pace rapidly over the past six months:

The company closed a $500 million Series C round at the end of last year, raised over $700 million in February this year, and completed another roughly $2 billion funding round in May.

Just one month later, reports surfaced that the company is seeking up to $2 billion in new funding, targeting an implied valuation of $30-31.5 billion, and is exploring a potential future public listing on the Hong Kong Stock Exchange.

For the capital market, the positioning as a global cutting-edge model developer is the core logic that justifies such a high valuation.

By launching K3, which it frames as "the world's largest model approaching the frontier of AI capabilities," Moonshot AI is not just putting on a technology showcase—it is presenting tangible proof to back up its valuation claims.

It needs to convince investors that Moonshot AI has not fallen behind the competition, and still maintains a seat at the table of the world's top-tier AI model developers.

In all fairness, Kimi's iteration speed does place it in the first tier of domestic AI model developers.

Over the past year, it has rolled out new model releases almost every quarter, its long-context capabilities and Agent cluster features directly address major pain points for developers, and its founding team has built a strong public reputation for technical excellence.

But the pressing question now is: among a crowded field of competitors each with their own unique strengths, what exactly is Kimi's core, differentiated market positioning?

Is it pursuing a high-premium, high-end product strategy? Without a clear generational performance gap over competing models, why would users be willing to pay 2-3 times more for their services?

Is it pursuing a price-competitive strategy? Cutting prices would only widen the gap between inference costs and revenue, amplifying losses and putting far more pressure on compute infrastructure investment and future fundraising efforts.

Is it pursuing an open-source strategy to build a broader developer ecosystem? Releasing model weights publicly would directly siphon off API revenue through third-party private deployments, and the widespread adoption of open-source models would continuously push down the overall market price ceiling, making future price increases nearly impossible.

At its core, this is a fundamental conflict between the compelling fundraising narrative and the harsh realities of commercial operations.

To maintain its lofty valuation and successfully go public, the company must uphold the narrative of being a "global cutting-edge developer with continuous breakthroughs," which requires pouring endless capital into R&D;

But the more it spends on R&D, the larger its losses grow, the more unclear its path to profitability becomes, and the more fragile its valuation logic grows.

On top of that, the Chinese AI market has its own inherent price ceiling. Domestic model developers cannot charge sky-high prices for their flagship models the way OpenAI and Anthropic do. The closer domestic models get to matching U.S. flagship capabilities, the more obvious the cost-revenue inversion becomes.

No matter how glamorous K3's launch event was, it cannot hide the fundamental question that Moonshot AI still faces: how to build a sustainable, profitable business.

The Dilemma of Model Competition

The harsher reality is that maintaining a seat at the cutting-edge table is becoming exponentially more expensive.

Many observers have argued that after K3's release, the gap between Chinese and U.S. large language models has narrowed further. This assessment is only partially correct.

Based on benchmark rankings and user experience, the performance gap has indeed shrunk—but the reality is that China is catching up to the flagship models that the U.S. released months ago, not the next-generation models that U.S. developers are actively training right now.

At the very least, xAI has publicly announced that it is pursuing multiple R&D paths, including a project targeting 10 trillion parameters. Foreign media reports have also revealed that OpenAI and Anthropic are developing next-generation foundation models on the 10-trillion-parameter scale.

The direction is crystal clear: next-generation models are being developed with far larger parameter scales, longer training cycles, and exponentially higher compute budgets.

This is no longer a simple matter of buying a few more GPUs.

A 2.8-trillion-parameter sparse model like K3 can require tens of thousands of H100-equivalent chips for a single formal training run, with the full project costing tens of millions or even hundreds of millions of dollars;

A 10-trillion-parameter sparse model would likely require a cluster of 100,000+ GPUs and a budget exceeding hundreds of millions of dollars.

What burns through capital even faster is that training is a one-time cost, but inference operations generate ongoing expenses every single day.

The more a model engages in extended reasoning, runs continuous Agent tasks, and works for hours on end, every user API call keeps draining the company's resources on compute hardware, power consumption, and network bandwidth.

But what truly creates an insurmountable competitive gap is not just the money spent on training and inference—it is the ability to run five distinct R&D paths simultaneously, and accept that four of them can fail without threatening the company's survival.

Top-tier U.S. AI labs are backed by cloud giants including Microsoft, Amazon, Google, and Oracle.

The costs of chips, data centers, network infrastructure, and fundraising can be spread across their entire cloud business, enterprise software revenue streams, and deep access to capital markets.

If one model training run fails, they can pivot to a different approach and start over. If inference operations run at a loss after a model is launched, they can absorb those costs using revenue from existing cloud contracts and their broader ecosystem.

Independent model developers like Moonshot AI do not have that luxury.

The billions of dollars it has raised are enough to fund one large-scale project on the scale of K3, and even attempt one more aggressive 10-trillion-parameter sparse model training run.

But that capital is nowhere near enough to support training multiple different model architectures in parallel, subsidize low-cost API access over a long period of time, and still release 2-3 new generations of models every single year.

For independent model developers, every flagship model launch is essentially a high-stakes gamble where they put their entire valuation, cash reserves, and next round of funding on the line.

This creates a very contradictory, awkward situation:

Domestic teams have used far less compute and clever algorithm design to rapidly close the gap on most publicly visible performance benchmarks. But the remaining top-tier, frontier capabilities are a pure game of raw compute power and massive capital investment—there are no shortcuts left.

In an era where training and inference costs are skyrocketing exponentially, the number of players that can survive to the next round of competition will only keep shrinking.

K3 has kept Moonshot AI seated at the cutting-edge table—but what about the next round, and the rounds after that?

This article is sourced from the WeChat Official Account "World Model Workshop", authored by World Model Workshop, and published by 36Kr with authorized permission.