Post-training isn't the savior of AI applications

@Guo Yunxiao

For the past few months, the market has been starved of compelling narratives. Embodied AI has hit a slump, interactive games are posting weak numbers, the world model that ranked #1 globally sold for just over $8 billion, personal agent startups are struggling to gain traction, and RSI remains irrelevant to most people. The only thing still making ripples is borderline companion apps.

AI applications remain stuck in an awkward cycle.

If you were an AI application founder trying to raise your next round right now, what would you do?

Some companies have started pitching stories about pivoting from applications to model companies. After all, post-training looks a lot easier than starting from pre-training. And model companies are worth far more than application companies.

But this narrative quickly suffered two reversals. First, people dismissed it as just another money-grabbing story. Then they realized post-training actually does help application companies — at least with cost optimization.

Post-training won't save application companies, but it turns out they still need to do it.

"Everyone Should Do Post-Training"

To put it simply: pre-training is like reading books; post-training is like doing practice problems. Whatever you train on, you get better at.

Post-training truly entered application companies' field of vision in the first half of 2026. Before that, AI application companies had explored virtually every possibility but achieved little, while model companies kept climbing steeply. People gradually came to understand the fundamental difference between model companies and application companies.

So the chase after model companies successively migrated to video models, 3D models, world models, and so on. The fatal allure for investors was that these companies offered them a chance — at an affordable price — to make up for missing out on the last wave of language models. So far, none of these other foundation models have become the next language model, and application companies certainly haven't either.

But soon, as these stories piled up, everyone realized post-training wouldn't turn an application company into a model company. Foundation models excel at covering many general scenarios and becoming infrastructure; theoretically, an application company's post-trained model should be as specialized and business-specific as possible. Which brings us right back to it being only relevant to the application company's own business.

Investors quickly cooled on the idea, deeming it too nebulous a story.

Then came another turn. Beyond the well-known post-training cases of Harvey and Cursor, in September Genspark announced their post-training results in collaboration with Fireworks AI, based on MiniMax M3 — a directionally strengthened slide-making model called Gen-1 Slides. Official documentation stated that this model produced slides at quality comparable to Opus 5, but at 1/17th the price. Sources close to overseas post-training platforms told me this is just one of several post-training models Genspark is advancing.

Manus takes the view that users don't need to know what model drives a task behind the scenes, whether it's theirs or not. Word in the industry is that Manus started post-training early and already applies it at significant scale. In their recent Manus 2.0 release, they mentioned that their new Cascade architecture uses 23.2% fewer tokens than the previous system, cuts task completion time by 28.2%, and reduces operating costs by 32%. Some portion of these gains likely comes from post-training.

A brief note on video models. Because Seedance held the lead for so long, it wasn't until the releases of WAN 2.7, 3.0, and MiniMax H3 that video agent companies seriously began considering post-training. One major direction in video model post-training right now is generation speed — when generation speed approaches or exceeds playback speed, it unlocks scenarios like long-range real-time interaction. From what we understand, some video application companies remain divided on whether to conduct large-scale post-training.

But overall, a consensus is gradually forming that applications built on language model coding capabilities should do post-training. This isn't a story that will get you your next round of funding, but it does help run a company.

Top-tier models are still too expensive. People are increasingly recognizing that UE (unit economics) is a critical issue. You need to at least not lose money on a per-user basis to operate a product normally. This problem essentially boils down to how to use cheaper tokens.

It's looking more and more like manufacturing.

Model supply remains a seller's market. Even looking only at official models from model vendors, different product lines are already diverging between pushing intelligence ceilings and handling everyday tasks. So supply chain management still needs to happen — thorough evaluation, price comparison, routing, even building some in-house — all reasonable choices. And post-training is an important build-in-house component of that supply chain.

Of course, the above is just a general impression, and different parties understand post-training differently. Investors talk about massive-scale RL environments (which is all well and good), but some early-stage founders are actually thinking about parameter tuning or small-scale preference learning.

Not That Simple

There aren't many secrets to post-training at the technical level.

Current post-training methods have largely converged into several layers: small-scale parameter tuning, such as LoRA, which mainly gets models to mimic specific formats; SFT (supervised fine-tuning), which uses paired "input-ideal output" data to get base models to imitate, achievable with thousands of data points and thousands of dollars. These are now prerequisite steps for complete post-training.

Going further, there are preference methods like RLHF and DPO, which use judge models or humans to label which of two existing answers is better. Here, training compute costs don't increase noticeably, but human (especially expert) annotation is expensive. This is a method of practicing language arts problems.

Then there's RLVR. Looking at this method alone, it really just needs a question bank and clear verification rules — no demonstrations for the model required. The model makes extensive redundant attempts on each problem, scores each approach, and learns which reasoning paths are superior. The Agentic RL everyone is discussing right now is essentially RLVR on more complex multi-step tasks in simulated environments.

There's a fundamental cost structure shift here: data increasingly serves rollout scope, task distribution, and verifier construction, and as inference scale expands, costs rise by orders of magnitude.

From an application company's investment perspective, post-training is almost embarrassingly flexible. Anything from tens of dollars to tens of millions of dollars counts as post-training, with viable operational methods at every level.

The trouble is, many application companies didn't incorporate post-training into their roadmap from day one, so there's a lot of catching up to do. How to build your own eval, how to get verifiable results at critical junctures, how to turn data into trainable assets — many items get vaguely lumped together as "building the environment," and each requires some feel.

Whether in investment or difficulty, post-training isn't as simple as it looks. But when you consider the returns, there's no clear benchmark for comparison. All of this leads most Chinese startups to talk about post-training while actually watching from the sidelines, or dipping a toe in at most.

This is also reflected in the ecosystem. Post-train as a Service is gradually maturing in Silicon Valley — traditional cloud providers, Neoclouds, inference vendors, NeoLabs, their businesses all touch on this to varying degrees. Quite a few companies can already offer standardized or fully managed post-training services, and because of components like eval, fully packaged services even carry shades of consulting firms.

Among these post-training service providers' customers, not all are AI applications. Many traditional enterprises looking to use AI well are also purchasing services, with some even training models that are more internal-facing and not primarily used in their downstream customer-facing services. Meanwhile, post-training and self-deploying open-source models is a path that fits data compliance needs very well. Here there are both the usual enterprise service ecosystem issues, and what we mentioned before: sufficient cash and deep business understanding are prerequisites for effective post-training.

In China, this isn't yet a thriving ecosystem. Cloud providers, inference vendors, and many Chinese companies in similar ecological positions haven't actually launched post-training businesses — their services stop at fine-tuning. Among startups, Macaron released a series of post-trained models this summer, pivoting to a post-training platform under the name Mind Lab, offering the simplest post-training services like LoRA. Some founders have received their post-training quotes, roughly in the several hundred thousand RMB range.

One investor once considered incubating a post-training service provider, but researchers doing post-training at some model companies told him, "Going out to start a business doing this is both grueling and unprofitable." Especially compared to data vendors, post-training is a heavy-asset, heavy-operations, heavy-sales business, but application companies are still only at the stage of brewing demand — whether it will explode is anyone's guess.

Why is demand for post-training among application companies (or other Chinese companies) so much thunder with little rain? The root cause comes back to those two things: sufficient cash and deep business understanding.

What Counts as Good?

Something people haven't paid enough attention to: eval is one of the most critical components of post-training.

People working on eval tell me this work is like being a product manager in the AI era: making users' subjective experiences as objective and quantifiable as possible, with the higher goal of defining and leading user experience. This way, evaluation work becomes a combination of a task distribution and a scoring system.

Eval and RL work toward the same end.

From the model company's perspective, third-party evals are public targets. Some benchmarks become consensus because they identify problems where different models show gaps that are worth paying attention to. And benchmarks have half-lives because once released, the evaluation itself becomes a training target. Turn the benchmark into a training environment, have the model practice repeatedly, and you get score improvements on these general evaluations.

An application company's eval doesn't need much public value, but it answers a more fundamental question: What problem is our company actually solving? For us, what counts as doing this well? This set of tasks and scoring standards is, in fact, the complete distillation of the company's understanding of its own business.

This at least partially explains the relationship between eval and model capability improvement.

Define a good benchmark, turn it into a training environment, improve model capabilities through post-training, erase capability gaps between different models on that benchmark, then define a new benchmark... In this cycle, evaluation effectiveness becomes the driving force pulling models forward. This is why some believe that as post-training enters large-scale RL, defining good problems becomes one of the few places where humans continue to matter.

Meanwhile, eval isn't static. As post-training progresses, previously subpar capabilities get unlocked and integrated into the business. Eval's job is to continuously identify the next capability where models vary in performance and the business needs to deploy it. Measure the gap, then use post-training to raise that capability.

Building good eval is the foundation for improvement and the starting point of post-training. But returning to AI application founders' perspective, grasping the essence of the business in rapidly changing times may be necessary yet somewhat demanding. So from some investors' view, this is also a window to observe whether a founder is "first principles"-driven.

Post-training doesn't correspond to any single-layer problem; it's a strategic judgment.

Companies still need to serve their customers well. For AI application companies, post-training is just one technical means among many. Post-training can obviously improve some performance, but whether the investment pays off isn't necessarily a calculable equation.

Whether doing post-training can take a company to the next level is another question entirely.

Cover image: Rosa Bonheur, The Horse Fair, 1852–55, Metropolitan Museum of Art