K3 106, Grok 111: this AI leaderboard is starting to contradict itself

The bottom line

K3 106, Grok 111, DeepSeek V4 Pro 92 - This list looks like the AI version of IQ ranking, but what is really outrageous is: it does not let the model answer multiple-choice questions, but allows community contributors to burn their own credits and put models to work on real tasks at a computer.

When I first saw the screenshots in the group, my first reaction was: Has K3 already pinned Grok and DeepSeek to the ground?

Open the Distributed Radar again and see that things are immediately reversed: the list will be refreshed every 60 seconds; Kimi Code, ZCode, and Grok are still in public internal testing Harness; the so-called IQ is a conversion scale customized by the site, not human IQ; and 111 points in a certain gear does not mean that the entire model has won.

This is much more interesting than a poster saying "a certain model is the best in the world".

Because what it tests is not the model at the press conference, but: what shell the model comes with, how high the inference is, how much time and money is spent, and whether it can finish the job in the end.

A list that makes the “number one” in the AI circle a bit embarrassing.

In the screenshot provided by the user, the numbers are:

• K3: 106 IQ
• Grok 4.6: 97 IQ
• DeepSeek V4 Pro: 92 IQ
• DeepSeek V4 Flash: 80 IQ
• GLM-5.3: 77 IQ

But when I checked the DeepSWE dashboard again on August 20, 2026, another set of dynamic observations had appeared on the current page: K3 still showed 106 overall, Grok 4.6 overall showed 98; some Grok 4.6 gears reached 110 and 111. DeepSeek V4 Pro shows 92 in DSH and about 97 in max gear.

The numbers have changed, not because the website is wrong, but because this is a living list. New tasks, new samples, and new Harness will continue to push old results forward. The "throne" you captured today may be replaced in a few hours.

Therefore, this article does not write K3 as the permanent number one, nor does it write a certain high score of Grok as a comprehensive overtake. A more accurate statement is: These configurations perform well in the current public tests, but the samples, gears and test cases must be viewed together.

What does an IQ of 106 mean? Do not mistake it for human IQ

The IQ of distributed radar is the station’s own display scale. According to the page, test completion percentage is multiplied by 1.5 to convert to Radar IQ.

To convert, 106 IQ corresponds to approximately 70.7% task completion rate; 100% completion rate is approximately equal to 150 IQ.

It expresses: How much work has been completed by a certain model configuration in this set of tasks and scoring rules.

It is not an IQ in the psychometric sense, let alone saying that K3 has an “IQ of 106” or Grok has an “IQ of 111”. If you post the screenshot directly as "AI IQ exceeds that of humans", it will have slipped from an interesting benchmark to a headline-making one.

The most exciting reversal: Grok 111 cannot become a god

On the current dashboard, the high and medium gears of Grok 4.6 have had results higher than the overall score, with high even reaching 111. However, this type of results comes from the public internal beta Harness. The page is clearly marked as a tentative score. The sample is insufficient and does not participate in formal recommendations, rankings and curves.

That means 111 points is more like a clue worth following than a championship trophy.

Same goes for K3 + Kimi Code. The current total number of K3 is 106. The page displays different levels such as max, high, and low at the same time. The number of questions completed and the success rate are not the same. The model name is only half the answer, the other half is hidden in the Harness and reasoning stalls.

This is what traditional “model rankings” most easily miss.

What this website tests is not actually a model, but a complete set of workflows.

Distributed Radar currently views Codex, DSH, Kimi Code, ZCode and Grok separately. It also separates the inference gears such as max, xhigh, high, medium, and low.

In other words, it's not comparing the abstract "DeepSeek vs Grok", but rather:

Can a certain model + a certain set of Harness + a certain inference gear + a set of tools and context deliver results in real engineering tasks?

The same model may behave completely differently if replaced with a different shell. Harness is responsible for tool invocation, file reading and writing, context management, memory and agent loop; the model is responsible for generating judgments, but whether it can be implemented in the end depends on whether the entire vehicle is properly connected.

Therefore, the performance of DeepSeek V4 Pro under DSH cannot simply be regarded as a duel with another provider configuration. Their APIs, toolchains, sample sizes, and costs may all differ. The most valuable thing about this list is that it puts these differences in plain sight.

It does not ask "can you answer it?", but drops the task directly into a codebase

The important DeepSWE test right now is not a few multiple-choice math questions, but a software engineering task. Pages demonstrate tasks including real-world code issues such as FastAPI option handling, Bandit cross-process taint analysis, type binding, and data snapshot differences.

The model reads the project, understands the context, modifies the files, runs the tests, and submits the patch. The final score is whether the patch can pass verification in a clean environment.

This is not the same species as "please write a bubble sort".

Will the ordinary Benchmark test model give a good answer? This kind of test asks:

Can you remember what you changed half an hour later without accidentally breaking another part of the project?

The most cyberpunk part: Netizens burn their own credits to score points for the entire network

Distributed radar does not mean that the webmaster rents a bunch of machines and tests them behind closed doors. Users can log in with GitHub, select their own Codex subscription level, claim tasks, give the prompts to Codex or the corresponding Agent for execution, and then contribute the results back to the public dashboard.

The site calls this action "riding on the radar" and the contributor ranking list is called the "radar ladder". The page currently shows a total pedal speed of approximately 135,419 U.S. dollars, a cumulative total of approximately 329.6 billion words, and approximately 770 contributors.

This means that behind the list is not a static promotional picture, but a group of people using their subscriptions to conduct public experiments: some people run one question, and some run dozens of questions in a row, and the results slowly converge into a map of models, Harness, time and cost.

Of course, this also means that data is not the standard answer to a one-time lock in the lab. The distribution of user-contributed tasks, account quota, network environment and execution strategy will all affect the sample.

If the client says it is successful, it does not count, and the server has to re-verify.

If users just clicked "success", the list would have turned into a large self-comfort scene.

The design of distributed radar is: after the client submits a patch, the server puts it back into a clean environment for re-verification; the task is run in a one-time isolated Docker container, and trace auditing and sensitive information inspection are performed at the same time. The site also clearly states that the client's self-reported success will not directly become the final score, and problematic submissions will be frozen and reviewed.

This makes it much more serious than "community contributors vote who is the strongest", but it is still not the same as a scientific experiment that has undergone complete peer review. Task distribution, sample bias, public internal testing and configuration differences of different Harness should all appear in the reader's judgment.

Why is Pompeii murals still included in the list? AI can’t just write code

Another channel of the distributed radar tested the adjacency recovery of Pompeii mural fragments: the model saw the broken mural fragments, judged which fragments were originally adjacent, and then used Precision, Recall and F1 scores.

This is not to shoehorn visual questions into the code list, but to observe another ability separately: spatial relationships, fragment matching, and reasoning under incomplete information.

The main site further attempts to put software engineering and visual spatial reasoning on the same radar with Composite Intelligence. This direction is not yet a mature "comprehensive intelligence ranking", but it reveals a very important judgment: Future Agent competition will not only occur in the code completion window.

What five things should readers check on this leaderboard?

First, look at completion percentage or F1, don’t just look at the big IQ numbers.

Second, look at Harness. Codex, DSH, Kimi Code and Grok are not the same execution environment.

Third, look at reasoning effort. max may be more stable, or it may just be slower and more expensive.

Fourth, look at the number of samples and whether it is tested internally. 7/8 and 174/269, the credibility is not the same magnitude; the tentative scores of the public internal beta cannot be used to make a final conclusion.

Fifth, calculate time and cost together. For real work, “3 points better success rate but 6 times more money” is not necessarily a win.

I think this ranking is much more practical than "who is the smartest": Give it a dollar and half an hour, how much work can it do for you?

Our judgment: The most valuable thing to follow is not K3, but this community-built test rig

K3’s 106 is very eye-catching, Grok’s 111 is more suitable for communication screenshots, and DeepSeek V4 Pro’s 92 allows people to see the real-life difference between the model and Harness combination. But what this article really recommends is not a specific champion, but the Codex Radar route.

Manufacturers will select the best-looking Benchmark at the press conference; private lists are at least willing to list failures, time-consuming, expenses and internal testing labels. It may not always be correct, but it is closer to the question we really want to know: Can this AI get things done in my project?

If you open the distributed radar today and see K3 106, Grok 111 or DeepSeek 92, please don’t rush to repost “a certain model is a god”. Note down the time, harness, stall, number of samples and cost and come back to it a day later.

Because the most realistic part of this list is:

Today’s king may be replaced in an hour; what truly lasts is when you finally start to see how the model works, rather than just seeing how it boasts at the press conference.

What to watch

• When will Kimi Code, ZCode and Grok end public beta testing and incorporate results into official recommendations?
• Whether the same model can accumulate large enough comparable samples under different Harness and inference gears.
• Can DeepSWE and Pompeii Fresco channels form a more stable cross-task capability portrait.
• Whether cost, elapsed time and success rate become the main criteria ordinary users use to choose an agent.
• After the scale of community contribution increases, can server-side review and task isolation continue to remain credible.

FAQ

Is Codex Radar’s IQ the human IQ?

no. It is a display indicator converted by the website based on the percentage of test completion and cannot be used for direct comparison with human psychometric IQ.

Does K3 106 mean it is already number one?

cannot. The list changes in real time, and Kimi Code is currently in public beta testing, and the tentative scores will not be included in the official recommendation. 106 can only show that the current public observation performance is outstanding.

Is the Grok 4.6 111 stronger than the K3?

No direct conclusion can be drawn. 111 is the dynamic observation value of a certain inference gear. The sample, Harness and internal testing status must be viewed together.

Where does distributed radar data come from?

From the results submitted by users after claiming tasks with their own subscription quota or API balance, the server will re-verify the patch in a clean environment and display the time, cost and completion status.

Is this list suitable for selecting production tools?

It is suitable as a clue and should not be the only basis. It’s best to do small-scale testing with your own codebase, budget, and task types.

Where is the website link?

• Distributed radar: https://deng.codexradar.com/?benchmark=deepswe
• Codex Radar main website: https://codexradar.com/en/

*Note: The data in this article comes from observations of public pages on August 20, 2026 and screenshots provided by users. Codex Radar is a dynamic community public testing platform, and the results will continue to change; Kimi Code, ZCode, and Grok are currently in public internal testing, and the tentative scores do not participate in official recommendations, rankings, and curves. *