Minimax M3.1 Flash (Fully Tested): Okay, this MODEL is PRETTY GOOD!

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: Minimax M3.1 Flash (Fully Tested): Okay, this MODEL is PRETTY GOOD! · Channel: AICodeKing · 15:15 · Watch on YouTube

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

Summary

The video tests Minimax M3.1 Flash on eight Kingbench 3 tasks, where it scores 53 out of 80 (66.25%), a substantial improvement over the earlier M3 entry’s 31.25%. Its strongest results are the folding table, contact lens case, and math problem, while several simulations and 3D elements have interaction or geometry bugs. The local panda-model training project includes real data, training artifacts, and a working interface, but the reviewer flags limitations in its validation and fact reliability. The reviewer considers the result promising for prototypes but recommends checking outputs and testing interactions, noting that pricing and generation speed for this preview were not verified.

Key points

M3.1 Flash scores 66.25% on Kingbench 3, compared with 31.25% for the earlier M3 entry.
The eight tasks total 80 points; the reviewer cautions that the percentage does not mean the model fully solved that share of tasks, since partial credit is possible.
The elevator simulation scores 3/10 because a code error prevents passenger spawning and the central interaction does not work.
The contact lens case earns 8/10, with working cap controls but solid-looking interiors instead of recessed compartments.
The folding table earns 9/10 for its convincing appearance and smooth, functional folding animation.
The archery game scores 4/10: target faces are misplaced, the fourth target has an animation error, and the timer pauses between shots.
The math task receives 10/10, as do all five models in the comparison sheet.
The local panda-training project scores 7/10; it includes training artifacts and a working interface, but its validation is narrow and its data contains a contradiction.
The wristwatch scores 5/10: timekeeping and customization work, but the hour markers cluster near the top and other presentation issues remain.
The reviewer says to verify availability and test the main interactions; published pricing and measured generation speed for this exact preview were not verified.

Chapters

Start	Chapter	Summary
00:02	Model overview and benchmark context	The reviewer introduces M3.1 Flash, explains the Kingbench 3 score, and distinguishes verified information about the preview from details that remain unknown.
02:09	Elevator, contact lens case, and folding table	The first three tests range from a broken elevator interaction to a polished contact lens case and a strong, smoothly animated folding table.
05:55	Panda SVG, archery game, and math task	The panda illustration is appealing but imperfect, the archery game has major gameplay bugs, and the math answer receives full credit.
08:41	Local panda training and wristwatch	The reviewer examines the training pipeline’s real artifacts and data limitations, then tests a customizable watch whose timekeeping works despite a flawed dial.
11:56	Scores, leaderboard, and final assessment	The video compares the final benchmark scores and concludes that M3.1 Flash is an encouraging improvement but still needs consistency and careful user testing.

Quotes

That’s a jump of 35 percentage points. — 00:02

This gets 9 out of 10, tying Grock 4.7 for the best score on this task. — 05:21

It doesn’t mean an independent source checked the statement. — 09:49

These are issues that a quick browser check can reveal. — 14:44

Overall, it’s pretty cool. — 15:15

Concepts and entities

MiniMax M3.1 Flash, KingBench 3, MiniMax Code, MiniMax M3, Gemma 2 2B, LoRA, fine-tuning, AICodeKing

Full transcript (original language)

00:02 [music] Hi, welcome to another video. So, today we’re testing Miniax M3.1 Flash. And this one has a pretty interesting result. It scores 66.25% on Kingbench 3 compared with 31.25% for the earlier Miniax M3 entry. That’s a jump of 35 percentage points. But once you open the actual generations, the picture gets more complicated. The folding table is one of its strongest results. The contact lens case looks good and its caps open. Then you get an

00:34 elevator simulation that breaks when you press the main button and an archery game with its targets in the wrong place. So, I went through the saved code, opened the previews, and tested the main interactions. We’re going to look at all eight tests, the scores, and where this model sits against the rest of the leaderboard. The official name is M3.1 Flash preview, and Miniax announced it for Miniax Code on September 27th. They’re presenting it as a text model for everyday development from bug fixes to complete features. The announcement

01:05 also shows five reasoning settings from low up to max. I couldn’t verify a published numerical benchmark table, public API pricing, or downloadable weights for this preview yet. The public API documentation still lists the older M3, so I wouldn’t transfer that model’s context window, multimodal features, or benchmark scores to M3.1 Flash. The numbers we’re testing here come from Kingbench, and the sheet doesn’t specify which reasoning setting was used for M3.1 Flash. For pricing context, Miniaax’s token plan page lists plus at

01:36 $20 a month, max at $50, and Ultra at $120. Those are platform subscription prices. They don’t establish the per token cost or exact access conditions for this particular model. So, check the model available in your account before paying for it. Now, KingBench 3 has eight tasks here, each scored out of 10 for a total of 80 points. There are interactive simulations, 3D objects, and SVG illustration, a game, a math problem, and a local model training task. That percentage is the share of available

02:09 points. It doesn’t mean the model completely solved 66% of the tasks because several results receive partial credit. And I’m reviewing the saved outputs here. So, this doesn’t establish how many attempts or corrections each generation needed. So, let’s get right into it. The first test is the elevator simulation. The request is for three elevators with each elevator carrying just one person. People should spawn on different floors, have their own destinations, and wait for another elevator. If the first one is occupied, hovering over a person should show their destination. Now, visually, we get a

02:40 dark interface with three separate shafts and eight floors. There are sliders for the number of people and the speed, a spawn people button and counters for waiting, boarded, and delivered passengers. But if we actually click spawn people, nothing useful happens. The counters stay at zero and the browser reports an error before a passenger appears. I check the code as well. It creates a number for the passenger’s horizontal position, then tries to call that number like a function that breaks the spawning process immediately. The elevator graphics also sit clipped down at the

03:12 bottom of the shafts in the preview. So the controls and the layout are there, but the central interaction doesn’t work. I can’t give it credit for working cues, destination tool tips, or passenger delivery just because there is code attempting those features. This gets three out of 10. For comparison, SEU 2 and Opus 5.5 get eight here, while Groke 4.7 and GPT6 sole get seven. This is the weakest result for M3.1 flash in the set. Now, the second test is a 3D contact lens case. It needs clearly marked left and right compartments, and

03:46 clicking the caps should open them. This is much better. We get a rounded white case, an orange accent for the left side, a blue accent for the right, and large L&R labels. The lighting and shadows give it a clean product render appearance. Clicking directly on the left cap opens it, and the status changes to show that the left side is open. There are also separate controls for each cap and a both button. I tested that too and both lids open. There is one visible modeling problem though. When the lids open, the insides look like solid white surfaces rather than proper recessed compartments. Looking at

04:19 the code explains it. The main body is solid and the attempted liquid and lens geometry sits underneath its opaque surface. So, I like the presentation and the working cap interaction, but the inside needs more work. This gets 8 out of 10. That matches SEU2 beats GPT6 Souls 6 and sits below Grock 4.7 at 9 and Opus 5.5 at 10. The third test is the folding table. The instruction is simple. Moving a slider to the right should unfold it. Moving left should

04:49 fold it. And the movement should be smoothly animated in 3D. And this is probably my favorite visual generation from this run. It has a light wooden tabletop with visible grain, rounded edges, metal legs, diagonal supports, and black rubber feet. It looks like an actual folding table. Moving the slider left brings the two leg frames underneath the tabletop, and the table lowers toward the ground. Moving it back to the right brings the legs out and raises it again. I checked the fully folded position, the halfway position, and the fully open position. The code also does something sensible here. The

05:21 two leg frames have different widths, so they can nest underneath the top, and the motion eases toward the selected position. That makes the change feel much better than simply snapping from one pose to another. This gets 9 out of 10, tying Grock 4.7 for the best score on this task. SUI 2 and Opus 5.5 get 8 and GPT6 soul gets seven. So, M3.1 Flash absolutely can produce a strong result when the geometry and interaction come together. Next, we have an SVG of a panda eating a burger. This one is cute. There is a rounded pastel background, a

05:55 panda with large black eye patches, pink cheeks, and paws holding a tilted burger. The burger has sesame seeds and separate layers for the lettuce, tomato, cheese, and patty. There are crumbs near the mouth and little decorative stars around the illustration. The weak part is the bite. There is a round piece of bun sitting near the mouth with an awkward cutout underneath it, so it doesn’t read as a clean bite taken from the burger. You can understand the intended scene, but the overlap between the mouth and the food could be much better. This is a static SVG, and that’s fine because animation wasn’t required.

06:29 It gets 7 out of 10. Sui 2 gets eight, Groke 47 gets four, and both GPT6 soul and Opus 5.5 get 10. Now, let’s come to the archery game. The requirement is four targets with faster completion moving the player up a leaderboard. At first, the presentation looks promising. It creates a landscape with mountains, a bright sun, grass, an archer, and a heads up display showing the timer, wind, targets, and arrows. There is a draw and release interaction and a dotted guide for the arrows trajectory. But when I start the game, the target

07:03 stands appear across the range, while the actual target faces are piled up in the top left corner. So, the things you’re supposed to aim at aren’t displayed where the game checks for hits. The code confirms the problem. It draws the target circles around the canvas origin without first moving the drawing position to each target’s location. There’s also a separate error in the moving forth target. Its animation uses a starting height that was never defined. And the timer has a problem, too. After I fired an arrow, it stopped at 1.24 seconds while I was preparing the next shot. The timer pauses whenever you’re aiming between shots, so it leaves out part of the time

07:35 you’re actually playing. That matters because this is supposed to be a time trial. A leaderboard based on that timer wouldn’t measure how quickly you actually completed the round. And the broken fourth target prevents a normal four target finish. Anyway, so nice scenery, but major gameplay issues. This gets four out of 10. Sway 2 and Grock 4.7 gets seven, GPT6, Soul gets six, and Opus 5.5 gets nine. The sixth test is the math problem. We have 60 ordered pairs arranged across 20 possible x values and three possible y-v values

08:08 with specific rules for how consecutive pairs can connect. We need to count the valid orderings starting on the top row and ending on the bottom row. The expected answer in the benchmark is 20,460 and m3.1 flash is recorded as correct with 10 out of 10. All five models in the comparison sheet get 10 on this one. So this is a full score result but it doesn’t separate these models. Now the seventh test is much more ambitious. It asks the model to create a data set of panda facts, fine-tune a Gemma 2 billion

08:41 parameter model and build a local web interface that produces a fact whenever the page refreshes. And there is substantial work here. The project contains 184 distinct facts expanded into 920 training examples and 184 validation examples. It uses a 4-bit Gemma 22B model with Laura adapters which are a smaller set of trainable weights added to the base model. There are actual saved model files and adapter weights. The training log reaches 600 iterations and the reported validation loss drops from 3.893 to 0.207.

09:17 So this isn’t just a page with a random sentence and a claim that training happened. The interface is called aopota. It has a warm paper colored background, large serif text for the fact, a category underneath, and another fact button. I opened the local app, clicked that button, and refreshed the page. It did produce different entries, but there are important limitations. The validation examples use the same facts as training with different prompts. So, that lower loss shows improvement on that narrow data set. It doesn’t prove the model can reliably answer new Panda questions. Also, the app compares

09:49 generated text with its own data set and can replace a close match with the stored wording. When it says verified, it means matched to that corpus. It doesn’t mean an independent source checked the statement. And I actually saw why that distinction matters. One displayed entry said the panda’s digestive system is built for bamboo rather than flesh. Smithsonian’s National Zoo describes it as more similar to a carnivore’s digestive system. And another entry in this project’s own data set says essentially the same thing. So the corpus contains a contradiction. There is a lot to like

10:21 about the local pipeline, but the data quality and reliability need work. This gets 7 out of 10, while the other four models in the sheet get 10. Finally, we have the 3D wristwatch. This needs to show the current time with smooth hands, included day and date, and track two time zones. The result is called Meridian GMT, and it is quite ambitious. It has a metal case, a strap wrapped around a display wrist, a crown, a dark dial, day and date windows, and a separate hand for the second time zone. There are controls for the dial color,

10:52 strap, hand movement, lighting, and camera angle. The live clock works in the preview. I changed the second time zone from London to Tokyo, and the readout updated from 5 hours ahead of New York to 13 hours ahead. The hand animation also uses fractional seconds, which is what you want for smooth movement, but the face has a very obvious geometry problem. The hour markers are all bunched near the top, forming this strange star-shaped cluster instead of being spread around the dial. The code gives them the same starting position and rotates each marker locally rather than placing it around the center

11:25 of the watch. The outer numbers also appear upside down. And clicking the dial camera preset brings us to a steep almost edge on view that makes the face harder to inspect when the button should help you see it. So the timekeeping and customization are there, but the actual watch face needs fixing. This gets 5 out of 10. SWE2 gets eight, Grock 4.7 gets six, and GPT6 Soul and Opus 5.5 both get 10. Now let’s look at the final scores. M3.1 flash gets three for the elevator,

11:56 eight for the contact lens case, nine for the folding table, seven for the Panda SVG, four for archery, 10 for math, seven for the local training task, and five for the watch. That adds up to 53 out of 80 or 66.25%. In the detailed comparison, Gro 4.7 gets 62 out of 80 or 77.5%. GPT6 Soul at medium gets 66 or 82.5%. Sweat 2 gets 67 or 83.75%. And Opus 5.5 at medium gets 75 which is 93.75%. The separate leaderboard graphic

12:29 says 93.5 for opus but 75 out of 80 gives us 93.75. So I’m using the detailed sheet for that comparison. Looking at the full leaderboard M3.1 flash sits just above croc 4.5 at 65 and below Q3.8 flash next at 70. Above those GPT 5.6 Soul and Muse Spark 1.3 both have 71.25. 25. DSV4 flash has 72.5 and GLM 5.2 has 75. Muse Spark 1.2 and DSV4 Pro both have 76.25.

13:01 Kimmy K3, Opus 5 and Grock 4.7 are tied at 77.5. GLM 5.3 Flash has 78.75 and Opus 4.8 has 80, Quen 3.8 max, Gemini 3.8 Flash and Deepseek V4.1 Flash all have 81.25. 25 GPT6 Soul and Fable 5 have 82.5 followed by SWE2 at 83.75 and the 3.8 MAX0902 entry at 86.25 at the top GPT6 Astra has 90, GLM 5.3 has 91.25, Fable 5.1 has 92.5 and then we

13:37 have Opus 5.5 with the percentage discrepancy I just mentioned. Below M3.1 flash we have Muse Spark 1.1 at 60, GPT 5.6 Terra at 56.25, 25, Opus 4.7 at 50, Sonnet 5 at 48.75, GPT 5.6 Luna at 40, and GPT 5.5 at 35. The earlier Miniax M3 entry has 31.25, Gemini 3.5, Flash has 30. The separately labeled Deepseek V4 Pro entry has 24.8, and High3 has 20. Those are the labels and values shown in the supplied chart. In particular, it lists Deepseek V4 Pro and DS V4 Pro

14:11 separately, so I wouldn’t silently combine those entries without knowing the difference between the runs. So, the big improvement over Miniax M3 is clear in this benchmark, but M3.1 Flash still trails the stronger models overall, and its results vary a lot depending on the task. Personally, I like the folding table and the contact lens case gives us a useful starting point. The local training project also shows that it can put together several real components, including data preparation, training, inference, and a web interface. The problem is that some basic mistakes survive into the finished output. A

14:44 button crashes before spawning anything. Target graphics are drawn in the wrong place. Watch markers all share the same position. These are issues that a quick browser check can reveal. So, if you’re trying it for prototypes, inspect the result and test the main interaction before building on top of it. The score improvement is encouraging, but I would still want better consistency before trusting it with less supervision. And without verified pricing for this exact model or measured generation speed, I wouldn’t declare it the best value just

15:15 because it has flash in the name. Overall, it’s pretty cool. Anyway, let me know your thoughts in the comments. If you like this video, [music] consider donating through the super thanks option or becoming a member by clicking the join button. Also, [music] give this video a thumbs up and subscribe to my channel. I’ll see you in the next one. [music] Until then, bye.