Has David toppled Goliath?

The opinion of most AI researchers is that frontierHighly capable, general-purpose AI models at the leading edge of what current systems can do. open-weight modelsModels whose trained parameters can be downloaded, run and modified. This does not necessarily mean their source code, training data or training process are also public. lag behind the most capable models by an average of 4-6 months. Following the release of Kimi K3 and, three days later, Qwen 3.8, we think it's fair to say that the gap has narrowed. This is our review of the two models and what these new releases mean for the industry.

Kimi K3

On the 16th of July 2026, Moonshot released Kimi K3, 371 days after the release of Kimi K2, a model that impressed the community with its ability to generate decently sufficient code at a low price point. We at Closp also enjoyed Kimi K2; it was a cheap alternative to Anthropic's models that you could plug into Claude Code. As the year progressed, Moonshot's presence became less and less known. They released Kimi K2.5, and although it enabled some nice models like Composer, it wasn't as impressive a step up as K2. But now Moonshot seems to have done it again and awed the community, this time with a much more impressive model that, in some domains, seems to beat out even the top dogs over in the US.

Kimi K3 is the first open-source 2.8T-parameter2.8 trillion parameters. Parameters are numerical values a model learns during training and uses to produce its outputs. model. It has vision capabilities and a 1-million-token context windowThe maximum number of tokens a model can handle in one interaction, including both the input and its output.. Served via the APIApplication Programming Interface: a way for one piece of software to communicate with and use another program or service. it has a throughput speed of 20 TPSTokens per second: a measure of how quickly the system generates output.. Pricing is $0.30/MTokMillion tokens: the unit AI labs commonly use to quote model prices. for cache-hit inputInput the system has already processed and stored for reuse, so it does not need to be processed from scratch., $3.00/MTok for cache-miss inputInput that is not already stored in the cache and therefore has to be processed from scratch., and $15.00/MTok for output.

Although its overall performance still trails that of the most powerful proprietary modelsModels controlled by their developers and generally accessed through the developers' own products or APIs rather than by downloading the model weights. (see below benchmarkStandardized tests used to compare how models perform on particular tasks or abilities. results), it is definitely a frontier model: the first open frontier model.

In some areas Kimi K3 even beats out the top closed models. The two most notable areas are browser usage and web development. Its web-development performance is evident from the scores on Arena AI: Kimi K3 is the top model in the Web Development section. Its strong browser performance is demonstrated by its BrowseComp scores.

Browser usage is particularly interesting: think about all the work tasks that require you to access some interface in a browser. Having an agentAn AI system that can plan and carry out a multi-step task, often by using tools, without needing continuous human guidance. that is able to navigate those interfaces is really useful for speeding up workflows and enabling agents to work more autonomously.

Furthermore, it shows some promising capabilities in 3D, which are fun but not very useful. Below is an example of a 3D Game Boy arcade emulator that the people over at Moonshot made using Kimi K3.

Moonshot's announcement also had a few examples that tried to hint that they are closer than the others to reaching some sort of snowball effect in AI development, where the models a lab develops lead to better and better models, since they can design chips, write GPU kernelsSmall programs that run directly on a graphics processing unit (GPU), usually to perform many calculations in parallel., help with training and so on. This is a classic Anthropic playbook. Anthropic loves talking about how their models will eventually create more powerful models that in turn create even more powerful models, and on they go until singularityA hypothetical point at which AI surpasses human intelligence and can improve itself, potentially causing extremely rapid technological change. is achieved. Announcements that hint at this effect should always be treated with a healthy amount of scepticism. But it is an interesting notion. In that vein, a few labs have started letting the models they are developing help with the later stages of development. Moonshot says that an early version of Kimi K3 handled most of the team’s kernel optimization during the later stages of development. OpenAI has made similar claims, saying that its models help diagnose failures, optimize training systems, run experiments and interpret results.

In summary, Kimi K3 is a good model; overall, it trails the top closed-source models but comes out on top in certain areas.

Qwen 3.8 Max preview

This review of Qwen 3.8 is going to be a bit different, and more touch and go than what you are used to from us. There won't be any token prices, TPS speeds or benchmark results. The model simply isn't fully out yet. It isn't available on the API, and the only way to access it is through Alibaba'stoken planAlibaba's subscription plan that bundles access to several AI models and charges usage against a pool of credits.. There are no official numbers on anything, and since it is only available via the token plan, benchmarkers haven't had the chance to test it.

All we have from Qwen is this tweet:

Qwen

@Alibaba_Qwen

·Follow

Qwen3.8 is launching and going open-weight soon!🌐 With a massive 2.4T parameters, this model is continuously evolving. We believe it’s one of the most powerful model available today, compatible to leading frontier AI models , second only to Fable 5. You don't have to wait to  Show more

8:29 AM · Jul 19, 2026

24.1K

Reply

Copy link

Read 1.3K replies

That's it in terms of official numbers. We know that it is a 2.4T-parameter model (smaller than Kimi K3) and that it is supposedly “second only to Fable 5.” in terms of capability.

So, due to the circumstances, we decided to conduct some of our own tests by pitting it against GPT 5.6 xHighThe model's second-highest reasoning-effort setting, which allows it to spend more tokens working through a problem before answering. in two different coding tasks.

Test one: the LoRA speedrun

For the first test we gave both models the same job: bolt a LoRALow-Rank Adaptation: a fine-tuning method that freezes the base model's weights and only trains small extra matrices attached to it. Much cheaper and faster than updating every weight. adapter onto a small Qwen2.5 model and beat a GSM8KGrade School Math 8K, a benchmark of roughly 8,500 grade-school-level math word problems. Each one takes multi-step arithmetic to solve, and answers are scored by exact match. speed-and-accuracy benchmark. The agents were given access to a Modal account and were instructed to not spend more than $15 on computeThe processing power and hardware resources used to run AI training or other computational work..

This is a challenge anyone can try. It’s called LoRA Speedrun. It has a public leaderboard, although we did not submit these results to that leaderboard.

The test had to run twice, since the first time around GPT peeked at Qwen's job.

On the rerun, Qwen read the instructions and got on with the task it had been given. It trained a valid adapter and ran it thrice: 61.0%, 60.0% and 60.6% against a 57% bar, at about 218 seconds a run. GPT, on the other hand, completely brushed over the instructions. You see, the GitHub page for the speedrun contains two different tracks. Both models were instructed to complete Track 1 and not Track 2. GPT completed Track 2.

So Qwen wins on disqualification. How very conclusive…

Test two: multi-agent orchestration

For the second test we gave both models a more “normal” coding task: build a production-ready dashboard that, once a day, would scrape the websites of every lunch restaurant within two kilometres, as the crow flies, of an address in Stockholm and display each one’s dagens lunchSwedish for 'today's lunch': a fixed-price weekday lunch special that most restaurants post daily, a bit of a national institution.. (Shoutout to the IBM office in Malmö for the fun idea.) The distance had to be worked out from real coordinates. It had to be a monorepoA single code repository containing multiple distinct projects, such as an application and its supporting packages., though they got to pick the stackThe combination of technologies used to build an application, such as its frontend, backend and database. themselves. We told both models to use subagentsWorker agents a lead model spins up to handle part of a task and report back the results., which made it possible for us to understand how good each model was at orchestrating a small team of agents. Both divided the work pretty well. Qwen spun up four workers in parallel; GPT delegated three narrower tracks and kept the core of the build for itself. Qwen did some stuff GPT didn’t. It spawned a review agent to check the work before calling it done, which caught a few bugs. And when one of its workers died mid-run on a quota error, the coordinator built the missing package itself.

The token counts came out near-identical: about 36.6 million for Qwen and 34.2 million for GPT. But Qwen came back with more to show for it. It managed to get ten restaurants with working scrapers compared to GPT’s two. GPT was much more dishonest about its work. It declared its dashboard production-ready when only two of eighteen restaurants were automated, and only walked that back once we pushed. We also think Qwen produced a much nicer design. Qwen’s version had a map, filters and a dashboard layout. By contrast, GPT’s design looked more like a menu than a dashboard. This was another failure, given that the brief specifically called for a dashboard.

Screencaps of both finished dashboards are below.

Qwen 3.8 Max

Front page

Map view

GPT 5.6

Front page

Below the fold

Takeaway from tests

Okay, these tests are obviously not conclusive. They are closer to an anecdote than a rigorous test or benchmark. Furthermore, the tests also focused mostly on coding, not the many other tasks people use AI for. GPT 5.6 seemed to be in an extra pitbull-esque mood during these two tests, which made its performance quite lacklustre and therefore harder to compare fairly. But with that said, it was entertaining to see Qwen work and we think it did a good job. It feels fast and has a way of tackling problems that feels much more mature compared with how Sol performed. In that sense, it’s much more akin to what you’d expect from Fable. Wise rather than wild. We think this will be a pretty good model once it’s fully released. As for the claim that it’s “second only to Fable 5”: sure, it might just be.

Verdict

Betteridge's law of headlinesAn adage that states: "Any headline that ends in a question mark can be answered by the word no." It is based on the assumption that if the publishers were confident that the answer was yes, they would have presented it as an assertion. applies. We think it’s safe to say that “the open labs” have not toppled the closed ones. What the narrowing gap means is a much bigger subject than we can cover here. The geopolitical aspectMost open-weight labs are Chinese. and cybersecurity aspectThese models are great at finding exploits and are not bound by any programme like Trusted Access for Cyber or Project Glasswing. will have to wait for separate notes. Here we want to focus on what it means for you as a user and customer.

The closer open models get to the frontier, the less AI depends on a handful of companies deciding what access should cost and how their models may be used. More capable choices force every lab to compete harder on price, speed and quality, which is plainly good for customers. They also let companies use a smaller or specialised model when that is enough, instead of paying frontier prices for every single task. Because the weights can be run and adapted elsewhere, users have more control over their data, their infrastructure and the knowledge they build on top of a model. Openness is not automatically safe, but it does allow more researchers to inspect failures, test safeguards and find problems that a closed provider may miss. The best version of the AI industry is therefore not one with a single winner, but one where useful models are plentiful, competition remains fierce and switching providers is possible.

May the gap continue to narrow.

Sources

• Moonshot AI — Kimi K3: Open Frontier Intelligence
• Artificial Analysis — Kimi K3 Intelligence, Performance and Price Analysis
• Arena AI — Code Arena WebDev leaderboard
• Moonshot AI — BrowseComp scores
• Microsoft — Open Weights and American AI Leadership