Mistral Large 4 (Le Chonk - Fully Tested): Is IT ACTUALLY The BEST OPEN MODEL!?

Video authored by the channel shown, not by IANIA; all rights belong to the creator. Transcript and summary generated automatically; the original-language content is preserved.

Original language: English

Original video: Mistral Large 4 (Le Chonk - Fully Tested): Is IT ACTUALLY The BEST OPEN MODEL!? · Channel: AICodeKing · 08:53 · Watch on YouTube

Topics: AI agents, AI evaluation and benchmarks, Software engineering with AI, AI models, Local and open-source models

Summary

The video tests Mistral Large 4, called “Lchonk” in the transcript, on eight coding tasks using Open Code and Open Router, judging the results that actually run rather than the model’s explanations. It succeeds on several visual projects and completes a local Gemma fine-tuning workflow, but two initial attempts produce no usable answer, and the watch requires a browser-tested repair. With a permission-related rerun counted and separate retries excluded, the score is 42 out of 80, or 5.25 out of 10 overall. The reviewer recommends it for contained visual prototypes while advising users to keep a browser and test loop close for longer tasks.

Key points

The test uses eight KingBench 3 prompts, each worth 10 points, in fresh folders; browser projects are checked in Chromium.
The elevator simulation works, including passenger queues, and scores 9.5 out of 10 despite minor reset and counter issues.
The contact lens case gets zero in the default run because no file is produced; a separate retry builds working lids but is excluded from the total.
The folding table has a working 3D folding slider and scores 8.5 out of 10, while the panda SVG scores 8 out of 10.
The bow and arrow game includes the requested systems but has tiny targets and other usability issues, earning 6 out of 10.
The counting task produces no final answer in the default session and scores zero.
After a permission setting is corrected, the model creates 510 training examples, trains a LoRA adapter for 400 steps, and serves changing panda facts; this task scores 10 out of 10.
The watch initially fails to load, then works after the model repairs its import paths, but visual issues limit the repaired result to 6 out of 10; it is excluded from the total.
The reviewer’s overall score is 42 out of 80, or 5.25 out of 10, and the recommendation is to test outputs carefully rather than treat this run as a universal ranking.

Chapters

Start	Chapter	Summary
00:03	Test setup and elevator simulation	The reviewer explains the eight-task browser-tested benchmark and begins with an elevator simulation that works well, with a few minor edge cases.
02:07	Contact case, folding table, and panda SVG	The default contact-case run produces no file, while the folding table and panda artwork deliver usable visual results.
04:44	Bow game and counting task	The bow game has functioning systems but usability problems, and the counting session ends without a final answer.
05:47	Fine-tuning workflow and wristwatch	After fixing a permission setting, the model completes local Gemma fine-tuning, while the watch needs an import-path repair and further visual work.
07:51	Overall score and verdict	The reviewer summarizes the mixed results, gives a score of 5.25 out of 10, and recommends keeping a browser and test loop close.

Quotes

I’m judging the delivered result, not just a convincing explanation of what the model intended to build. — 00:33

the key request, a continuously controlled three-dimensional fold, is there. — 03:09

That’s a strong result for a task with so many moving parts. — 06:17

The score with that permission issue corrected is 42 out of 80. — 07:20

I would absolutely try Lchonk for a contained visual prototype — 08:22

Concepts and entities

Mistral Large 4, OpenCode, OpenRouter, KingBench 3, Chromium, Gemma, LoRA, Three.js, AICodeKing

Full transcript (original language)

00:03 Hi, welcome to another video. So, Mistral Large 4, also known as Lonk, looks interesting on paper, but you know what? I really wanted to find out whether it can actually finish the kind of coding tasks I give to models on this channel. I already covered the announcement, the specifications, the pricing, and Mistl’s published benchmark claims on my AIC King channel, so I’m not going to repeat that video here. If you want the full release breakdown, please check it out and subscribe to AIC King as well. In this video, we’re going straight to the hands-on results. I

00:33 connected the Mistral Large 4 preview to Open Code through Open Router. For every task, I gave it a fresh folder and used the same eight KingBench 3 prompts I have used in earlier model tests. Each question is worth 10 points, and I’m judging the delivered result, not just a convincing explanation of what the model intended to build. For the browser projects, I opened the actual output in Chromium. I checked whether the page rendered and whether the main interaction worked. That distinction matters because a model can write a beautiful summary at the end of an agent session while the page itself does something else entirely. So, let’s start with the elevator simulation. The prompt

01:05 asks for three elevators with room for just one person in each car. I want to spawn people on different floors, see their destinations on hover, and watch anyone left behind board the next available elevator. Now, look at what Lch made. There is a clean building view, three colored elevator cars, a people count field, and a spawn button. Passengers have different destinations and the cars move with door animations. More importantly, the queue actually works. I reset the simulation, spawn five people, and let it run in the browser. The waiting and riding counters both return to zero with no browser errors. So, they really did get

01:37 delivered. I found one small edge case. Resetting while a car is already in the middle of a trip can leave a brief exit animation from the old trip in the new scene. and the waiting counter changes when someone is assigned a car slightly before they physically step inside. Neither breaks the main simulation, but they’re the reason I would give this 9 and 1/2 out of 10 instead of a perfect score. For a one file visual build, that’s a very good start. The second question is a three-dimensional contact lens case. It needs clear L&R compartments, caps that open when

02:07 clicked, and one HTML file. This is where the first real consistency problem appears. In the default open code run, the model spent its turn reasoning about the case, then ended without creating the file. I had nothing to open or click. So for this first pass benchmark, it gets zero. I am not grading its plan as if it were a finished product. I also tried a separate run with open code supported no reasoning setting. This time, Lechon did make the case. I opened it in the browser and clicked each cap. The left and right lids opened independently. They fly a little too far upward and partly out of the frame, but

02:38 the requested interaction is real. That tells me the model can build this task while the default run still deserves its zero for reliability. Now the folding table is a much better result. The page loads a proper 3JS scene with lighting, materials, shadows, orbit controls, and a slider. Dragging the slider all the way left gives me a compact folded table. Dragging it right unfolds the tabletop and legs. I opened both states in the browser and both rendered without page errors. There is a little visual polish missing in the middle of the animation. The feet can lift off the floor and the model looks a bit small on

03:09 the screen, but the key request, a continuously controlled three-dimensional fold, is there. I’m giving it 8 and a half out of 10. This is the kind of output that would make me keep experimenting with the model for visual prototypes. The panda question sounds simple, but it tells me whether the model can produce a finished visual asset rather than merely describe one. The prompt asks for an SVG of a panda eating a burger. Lechonk produced a valid SVG that looks clean when rendered. You can clearly see the panda, the burger, the paws, and the

03:41 open mouth. The little weakness is the action. The burger is held right against the mouth, but I don’t really see a convincing bite. So, the subject is right, and the artwork is cute, but the word eating is a little ambiguous. That’s 8 out of 10 from me. Then, we have the bow and arrow game. This one asks for four targets, a timer, and a leaderboard where the fastest player moves up. Lechon did create a complete game screen with four target types, a wind indicator, aeropysics, hit tracking, and a leaderboard saved in the browser. I launched it and it ran without a JavaScript error. So, there is

04:12 real work here. But when I actually look at the playing area, the targets are tiny. They’re only a few pixels wide at a normal desktop size. The target positions are fixed as well, which means the far target can move off the screen on a phone. The aiming guide also does not perfectly match the projectiles drag. So, yes, it has the requested systems, but it needs another pass before I would call it a genuinely enjoyable game. That’s 6 out of 10. Before we get to the final three questions, I also want to let you know what I have been doing for members. I’ve been working on a lot of membersonly videos, and I have just started a weekly

04:44 recap series there. I take the major tools and releases I cover here during the week and put the important parts into one summarized video of about 10 or 12 minutes. So, if you don’t have time to watch every upload, you can come to that recap and stay up to date without missing the main stuff. The series is available on all membership tiers, including bronze and up. I’m also working on exclusive tutorials for some of the tools, so stay tuned for those as well. If that sounds useful to you, click the join button on this channel. Now, let’s come back to Kingbench. The sixth question is a counting problem. The correct answer is 20,460.

05:15 What I wanted was either that answer with a convincing method or at least a substantive attempt I could check. In this default open code session, Lchonk again used its turn without delivering a final answer. So there is nothing to award here. 0 out of 10. This is exactly why I test what comes out of the agent, not just whether it seems to be thinking hard. The seventh task is the big one. Generate a data set of Panda facts, fine-tune Gemma 2 billion parameter locally, and make a web page that gives me a new fact on every refresh. My first open code setup blocked access to the

05:47 cacheed model. That was my permission setting. So I enabled it and sent the exact same prompt again in a fresh folder. And this time Lechon did the whole job. It made a data set of 510 examples, trained a Laura adapter for 400 steps, and brought the validation loss down from about 7.97 to 0.31. It even found and fixed a server threading issue during its own endto-end test. I opened the final page in Chromium and refreshed it three times. Three different Panda facts, no browser errors. Five more refresh requests also

06:17 gave me five different facts. The model and the web page both run locally, so this one gets the full 10 out of 10. That’s a strong result for a task with so many moving parts. The final question is a full three-dimensional wristwatch. It needs the current time smoothly moving second, minute, and hour hands and two time zones. The model wrote a fairly ambitious watch page and even downloaded the 3JS modules so they could be served locally. But the first browser check showed a blank page where the watch should be and both time displays still had placeholder dashes. I traced

06:49 that to the import map. The HTML requested the orbit and environment modules from paths that the model had not actually created. So first pass score zero because a watch that never loads is not a working watch. Then I gave open code the exact browser error and let the same model repair its own project. It fixed the paths. And on the second check, I got a 3D render, a clock that advanced, a sweeping second hand, and a second time zone that changed when selected. That repair is worth showing because it tells a more useful story than a screenshot of a broken page. But

07:20 I would still only give the repaired watch 6 out of 10. The wrist and bracelet dominate the frame. Parts of the case cut across the dial, and changing the second zone does not immediately update its printed label on the watch face. So, the time logic works. The visual execution needs more care. The score with that permission issue corrected is 42 out of 80. That’s 52.5%. Or 5.25 out of 10 overall. I’m keeping the contact case retry and the repaired watch outside that total because those were separate attempts after the model’s

07:51 initial output. The fine-tuning rerun does count. The first stop came for my open code permissions, and I don’t think it would be fair to score that against the model. What stands out to me is the contrast. When Lchon finishes a focused visual task, it can make something really nice. The elevator is genuinely functional. The folding table is polished and the Panda SVG looks good. And the local Gemma fine-tuning job shows it can complete a longer workflow when the environment is set up correctly. But the contact case and counting sessions still ended without

08:22 usable answers, and the watch needed a browser tested repair. Those are workflow problems you will feel if you’re relying on an agent to complete a task while you do something else. I also wouldn’t turn this single run into a universal model ranking. These are my prompts in this open code and open router setup with the results you can inspect on screen. I would absolutely try Lchonk for a contained visual prototype and I would keep a browser and a test loop close by for longer less forgiving work. I want to see it finish reliably before I trust it as my default coding model. For me, the deciding

08:53 factor is how often I can hand it a task and come back to something I can use. Overall, it’s pretty cool. Anyway, let me know your thoughts in the comments. If you like this video, consider donating through the super thanks option or becoming [music] a member by clicking the join button. Also, give this video a thumbs up and subscribe to my channel. I’ll see you in the next one. Until then, bye.