Open Models Are No Longer the Backup Option

Kimi K3 became one of my default models

I did not start using Kimi K3 because of a benchmark chart.

I started using it because I was working on my own product OrgWorkspace and needed a model that could handle a large amount of context without losing track of the actual objective.

It worked extremely well.

Not perfectly. No model does.

But it understood the product, stayed useful across complex tasks, and produced work that required far less correction than I expected.

I began using it more frequently.

Then I added it to Cursor.

Now, when I have a complicated task involving several files, a large codebase, planning, or a long sequence of changes, Kimi K3 is often one of the first models I try.

Sometimes I choose it before the latest models from OpenAI or Anthropic.

That still feels strange to write.

A few years ago, open models were interesting because of what they represented. You could inspect them, modify them, deploy them yourself, and avoid becoming completely dependent on one AI provider.

But if you wanted the most capable output, you usually returned to the closed models.

Kimi K3 does not feel like a philosophical choice.

It feels like a practical one.

Moonshot AI describes Kimi K3 as a 2.8-trillion-parameter, open-weight, multimodal model with a one-million-token context window. It was designed for long-running coding, knowledge work, reasoning, and agentic tasks.

Those numbers are impressive, but they are not why I keep using it.

I keep using it because it completes the work.

The gap has become difficult to see

People often discuss AI models as if there is a fixed hierarchy.

OpenAI and Anthropic are at the top. Google is somewhere alongside them. Chinese and open models are below, competing mostly on price.

That hierarchy is becoming outdated.

The new generation of open models may not win every benchmark or every task. Claude and GPT models are still excellent, and there are situations where I strongly prefer them.

But the gap is no longer obvious during normal work.

Kimi K3 can handle long coding sessions, research, tool use, and complex reasoning. Qwen continues to release capable models across coding, multimodal work, and general knowledge tasks. Z.ai has been improving the GLM family at a pace that is becoming difficult to ignore.

Alibaba’s newly released Qwen3.8-Flash-Next is particularly interesting because it focuses on efficiency, long-context performance, and reducing the cost of running agentic workloads. It is also an experimental preview of the architecture planned for Qwen4.

Z.ai has released GLM-5.3, focused heavily on complex software engineering and long-running agent tasks.

These are not small models being carefully compared against an older GPT release.

They are being designed for the same work people currently give to Codex, Claude Code, Cursor, and other serious AI development tools.

The competition has moved from “Can an open model produce a reasonable answer?” to “Which model completes this particular task most reliably?”

That is a much healthier question.

Ox Alpha proved something before anyone knew what it was

Ox Alpha made this shift even more obvious.

The model appeared on OpenRouter and OpenCode without a known creator. It was presented as a stealth model built for coding, reasoning, multimodal work, and long-running agentic tasks.

People started using it without knowing which company had trained it.

They only knew that it was surprisingly good.

There was plenty of speculation. Some people thought it came from a Chinese lab. Others tried to identify it through its writing style, tokenizer behavior, or the way it handled certain prompts.

Then, while I was preparing this article, the mystery was resolved.

Z.ai revealed that Ox Alpha was an anonymous preview of GLM-5.3-Flash.

That is fascinating because people were judging the model before they could judge the brand behind it.

There was no OpenAI logo.

There was no Anthropic announcement.

There was no keynote telling everyone how revolutionary the model was.

People tried it, gave it difficult work, and decided it was good.

Only afterward did they learn that it came from Z.ai.

That may be one of the clearest signs that the market is changing.

The model earned attention through its output before its creator earned attention through marketing.

The real advantage is not only price

Open models are usually discussed as cheaper alternatives.

Cost matters, especially when an AI product processes millions of tokens or runs several agents at once. A small difference in API pricing becomes significant very quickly.

But I think token efficiency needs to be understood more broadly.

The cheapest token is not always the cheapest task.

A model may have a lower price per million tokens but repeatedly misunderstand the request, rewrite the wrong files, or require three attempts to complete the work.

Another model may cost more per token but finish correctly on the first attempt.

The useful measure is not only how much each token costs.

It is how much useful work is completed for the total cost.

This is where the new open models are becoming genuinely competitive.

They are not simply producing inexpensive answers. They are remaining coherent across larger contexts, using tools, navigating repositories, and completing longer workflows with fewer interventions.

That makes them useful inside real products.

It also gives developers leverage.

If Kimi, Qwen, GLM, and future open models can perform the same task reliably, I do not have to design a product around a single provider.

I can route different tasks to different models.

I can use one model for planning, another for coding, and another for verification. I can switch providers when pricing changes. I can test new releases without rebuilding the entire product.

The model becomes a replaceable part of the system instead of the system itself.

“Open” does not mean you can run it on your laptop

There is an important reality hidden behind all this excitement.

People often hear “open-source AI model” and imagine downloading it onto a laptop, disconnecting from the internet, and having their own private version of ChatGPT.

For smaller models, that is increasingly possible.

For something like Kimi K3, I wish you the best of luck.

Kimi K3 has 2.8 trillion total parameters. Its Mixture-of-Experts architecture activates only part of the model for each token, which makes inference more efficient than activating the entire model every time.

But the rest of the weights do not magically disappear.

They still need to be stored and made available to the system. Running the complete model at a useful speed requires a level of memory, bandwidth, and GPU infrastructure that most individuals and smaller companies simply do not have.

Quantization, offloading, and improved inference engines can reduce the requirement. Enthusiasts will continue finding clever ways to run enormous models on increasingly unusual hardware setups.

That does not make local deployment practical for the average person.

This is why I prefer the term “open-weight” when it is more accurate.

The weights may be available. Developers may be able to inspect, modify, fine-tune, or host the model under its license.

That does not mean everyone can afford to operate it.

For most of us, the realistic way to use these models will still be through an API, a cloud platform, Cursor, OpenRouter, or another managed provider.

The model may be open.

The compute certainly is not free.

The product around the model matters even more now

A model is only one part of an AI product.

OpenAI has ChatGPT, Codex, computer use, connected applications, memory, research, image generation, and an interface that allows all these capabilities to work together.

Cursor offers something different that has become more valuable to me: choice.

I previously used Claude Code heavily, but I no longer do. The latest Opus 5 and Sonnet 5 models have not been good enough for the way I work, and Claude Code is naturally centered around Anthropic’s models.

Cursor gives me a much wider range of options.

I can use Kimi K3 for one task, switch to a GPT model for another, and test Grok, GLM, or whatever capable model appears next without abandoning the same development environment.

That flexibility matters more as the intelligence gap between models becomes smaller.

I may prefer Kimi K3 for a particular complex task, but Kimi does not need to provide the entire working environment. Cursor provides the environment, while I choose the model that best fits the work.

The winner will not necessarily be the company with the single smartest model.

It may be the company that makes several excellent models easiest to use together.

That is also part of the thinking behind Scalong OrgWorkspace, the product I am building. Instead of forcing every workflow through one model, OrgWorkspace can use different models and agents for different parts of the work, while keeping the overall process governed, auditable, and connected to the systems a business already uses.

Model loyalty makes less sense now

I used to think about AI tools largely through their flagship models.

Which company currently has the smartest model?

Which one writes the best code?

Which subscription should I keep?

I increasingly think that is the wrong way to look at it.

Different models are becoming good at different types of work, and the difference between first and fifth place is often smaller than the difference between using the right workflow and the wrong one.

Kimi K3 has become one of my favorite models because it works well for the tasks I actually give it.

That does not require me to declare that it is universally better than GPT or Claude.

It only needs to be better for enough of my work that I keep choosing it.

Qwen will improve.

GLM will improve.

OpenAI, Anthropic, and Google will respond.

Another model will appear next month, possibly under a strange anonymous name, and people will spend a week trying to identify who built it.

This competition is good for everyone building with AI.

Prices come down.

Context windows grow.

Models become more efficient.

Providers have to earn loyalty instead of assuming it.

And developers gain more freedom to choose the model that fits the work.

The gap is no longer the most interesting part

Open models have not defeated closed models.

They do not need to.

They have already changed the market by becoming credible alternatives for serious work.

A year ago, choosing an open model often meant accepting a visible compromise in capability.

Today, I can choose Kimi K3 for a complex task because I genuinely prefer its output.

That is the shift.

The question is no longer whether open models will catch up.

For many tasks, they already have.

The more interesting question is what OpenAI, Anthropic, and every other closed provider will build when model quality alone is no longer enough to keep us inside their ecosystem.

I still use ChatGPT.

I no longer use Claude Code.

I moved to Cursor because it gives me access to a much wider range of models without forcing me to change my entire workflow whenever a better one appears.

The latest Opus 5 and Sonnet 5 releases have not been good enough for my work. More importantly, I no longer want my development environment tied to whichever model one company happens to release next.

Inside Cursor, I can use Kimi K3, GPT, Fable, Grok, GLM, or whichever model gets the work done best.

I am no longer waiting for one company to release the next great model before my workflow improves.

There are now excellent models coming from everywhere.

And I am happy to use whichever one gets the work done.