Ling 3.1 Flash Hands-On: Free 560B Model vs Real Tasks

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: Ling 3.1 Flash Hands-On: Free 560B Model vs Real Tasks · Channel: Fahd Mirza · 09:53 · Watch on YouTube

Topics: AI agents, AI evaluation and benchmarks, Software engineering with AI, AI models, Health and science

Summary

The video tests Inclusion AI’s Ling 3.1 Flash, described as a hybrid reasoning model with around 560 billion total parameters, about 25 billion active per token, and up to a million tokens of context. In a Blender-and-Go game task, the result was largely broken, while a self-contained animated SVG city scene was judged much more successful. The model performed strongly on a SQL optimization task, identifying a costly correlated subquery and testing its proposed change; its clinical-assistant and multilingual responses also had strengths, though the latter faltered on some low-resource languages. The reviewer concludes that Ling 3.1 Flash is not best-in-class but can be a useful general-purpose daily driver.

Key points

The model is presented as designed for coding and agentic work, and as not open source at the time of the video, with a release said to be planned soon.
The benchmark chart covers 12 tests, including practical agent tasks, cybersecurity, coding, banking and finance, and healthcare.
The reviewer notes that the model is free and describes local use as inexpensive, while cautioning that it is not best of the breed.
The Blender-generated game had visible issues: the objects were disoriented and the game did not work as intended.
The animated SVG city scene was considered a much better result, though the reviewer wanted a denser, richer skyline.
On SQL optimization, the model identified a correlated subquery as the main performance bottleneck and tested a pre-aggregated join.
The reviewer praised the SQL response for explaining what not to change, as well as for testing its suggestion with synthetic data.
The clinical case response was described as prioritized and concise; the reviewer advised viewers to consult a real health professional for actual medical advice.
The model returned translations for 80 languages, with major languages judged mostly correct and some low-resource languages less successful.

Chapters

Start	Chapter	Summary
00:01	Introducing Ling 3.1 Flash and benchmarks	The reviewer introduces the model’s stated capabilities, starts a Blender-and-Go task, and discusses its benchmark results and low-cost positioning.
02:38	Testing Blender game creation and animated SVG	The generated game has major problems, while a separate prompt for an animated night-city SVG produces a more promising result.
04:44	SQL optimization test	The model identifies a costly correlated subquery, proposes a pre-aggregated join, and supports its recommendation with performance tests.
06:47	Clinical and multilingual tests; verdict	The reviewer evaluates a clinical-assistant response with a medical disclaimer, checks translations in 80 languages, and gives an overall daily-driver assessment.

Quotes

It’s a hybrid reasoning model with around 560 billion total parameters only about 25 billion active per token and up to a million tokens of context. — 00:01

these models start breaking apart as soon as we give them something complex. — 03:40

It also replaced it with one pre-agregated join — 06:16

10 out of 10 for this one — 06:47

I think it is uh quite a good daily driver. — 09:53

Concepts and entities

Ling 3.1 Flash, InclusionAI, Ant Group, Hermes Agent, Blender, Go, SQL, tool calling

Full transcript (original language)

00:01 We have a new link model from inclusion AI. Meet link 3.1 flash from ant groups inclusion AI. It’s a hybrid reasoning model with around 560 billion total parameters only about 25 billion active per token and up to a million tokens of context. At the moment it is not open source but they are saying they are going to be in the next week or so. It is built for coding and agentic work. It placed near the top of

00:33 the openware designed leaderboards which we will be talking about shortly. But for now, let me take you to my terminal where I already have configured it with Hermes agent. What I’m going to do, I’m going to give it this goal first where it needs to create its own 3D assets with Blender, then turn them into a working go dot game by itself using real tools with no handh holding. While it runs, let’s talk bit more around the benchmarks. And don’t worry about these warnings. I think Hermes

01:04 agent is losing its guest now. Uh there are a lot of issues with it. I keep saying that I’m going to switch which I will. Anyway, look at this benchmark which is bit blurry but this is what comes straight from ant group. So by closely looking at it the chat uh the the chart spans 12 tests covering real world work tasks banking and finance agents cyber security codebased questions and healthcare ling does best where agents have to do practical work.

01:35 It sits at the top or near the top on frontier suite cyber gym and healthbench professional and it holds its own agentic style task like skill bench and multi- challenge. It is not a clean sweep as you can see. Clot Opus 5 still etches it out on few panels like terminal bench and the small print says the health bench result came from their own AQ environment. So look I think at the moment the only story is the same which is the case with all of the lingering Bing u models from inclusion

02:07 AI. They are free. They are not best of the breed but they try to get the job done as a daily driver. the price is very low especially if you install it locally it is negligible just the cost of what you are doing and if you’re looking to host it locally you can rent the GPU from mass compute and I will drop the link to their website in videos description also recently I have started this free weekly AI newsletter which you can subscribe at vadmeza.substack.com substack.com and if you want to help out

02:38 the channel, please become a member. Okay, let’s go back and see it is still working through it. I’ll just allow it for this session. It is just running my terminal commands where blender and go do. It has produced these files after an hour. Let me check it out. I’m just going to run it in front of you. So, the game panel um in go loads but nothing appears. No, it is appearing. Just wait. Sorry, I was too quick to say. Okay. Anyway, I’m just going to press some

03:10 buttons here. So, something is happening. But you can see that trampoline is there, but it is off. I’ll just make it bit bigger. And I really don’t see um any landing here. So, land in the homep. I’m just pressing space. Let me move arrow keys. Okay. I’ll just come back here. Okay, left and right up and down they

03:40 move but they are totally disoriented here and you can see that it has tried to create some things but totally failed. So you can see that these models start breaking apart as soon as we give them something complex. Anyway, let’s try a few more. If you go through the information which ant group has shared, they claim that ling ranks near the top for SVG and mobile app design. So I’m testing that directly. One prompt, one shot and the

04:10 result opens in a a browser. I’m asking it to create a single self-contained HTML file with an animated SVG scene of a night city skyline scene from rooftop. And then there are few other things. So let me run that script. Let’s wait for it to generate something and we will check it out. And it has come back with the script. As you can see that completion tokens were just over 20,000. Let me open this thing in front of you. I’ll just drag and drop.

04:44 Pretty good. Pretty good. Very nice. So, it’s a night scope and there is this blinking tower. Some other skyline. uh not as dense or as um rich I would say but still not bad at all unlike the first test I think this one is much much better okay now let’s proceed to the next test where we are looking if link 3.1 flash can look at a slow tangled SQL query and

05:15 pick the one change that makes the biggest speed difference without listing 10 fixes or rewriting everything so Not only I’m checking its instruction following but if it is able to understand the SQL optimization the query is fine and I’m not asking it to just do 10,000 things just give me one thing for the greatest bang for the uh buck. So you see that I have just given it this prompt. Let’s see how it goes with it. I’m using the harness again

05:46 with this because I wanted to iterate on this. And one thing I have to say it tool calling is quite good because I gave it the file. It was very um it was able to smoothly access that file using its tools and then read it and now going through that query and checking every line out. Okay, being a former DBA I can tell you the model has done really good here. It picked the right target, the correlated

06:16 subquery that recomputes each opponent’s win rate per row. It also replaced it with one pre-agregated join and also put in a filter, moved into the join and also told us what is equivalent. So it didn’t stop at theory. It built a synthetic data set, timed it, and ran an ablation showing that subquery is nearly all of the runtime. So really good stuff. So I think uh one change no list no full rewrite it also explain why

06:47 indexes or rewriting the outer query wouldn’t help which is the strongest part of its response that what not to do really good 10 out of 10 for this one okay and now let’s test out another claim which they have made in their model card they say that it can act as a clinical assistant so I’m testing if it can do that if it can take a realistic emergency case and give a safe, correct, prioritized consult note. But before I

07:18 do anything, um, make sure you understand that this model is not a clinician. So if you want a real actual advice information, go consult a real health professional. This is not an alternative at all. Okay. So this is the prompt in your screen and this is the answer which model has given. So it got it right in about 22 seconds. By the way, I I have no clue what exactly this mumbo jumbo means, but

07:48 I have just checked it with a judge model. So what it has done, it has called an inferior heart attack, then caught the trap in the case. The low blood pressure points to a right, some um infrared, so it’s set to check a V 4R lead. So all in all it has kept to the brief with five clean sections and no rambling at all which is quite good. I’m quite impressed by this model’s instruction following. So look all in all as I said earlier this is not you

08:19 know best of the breed like fable or something like that but it is not bad at all and they are improving fairly rapidly. So link models are you know generally they are general purpose as you can see. So finally just to check the language and stuff I’m going to do a multilingual test to wrap up the video. So for this test we are asking link 3 flash to say will you marry me in 80 languages in a single request from

08:49 English to rare ones like ferroise baloi and triia and few others and also you can see we have gibberish and few others and this is the response. So I will scroll through please check your language and tell me if it is right or not. It has returned all 80 languages by the way in about couple of minutes which is good. The thing is that the major languages took you know all look right to me including Spanish, French, German, Russian, Mandarin, Turkish and few others. It also gave the Arabic in the

09:21 female form which is correct. The rare and low resource languages are where it breaks down. I don’t think so it has done a good job like Tamil u havan yuruba gutnesh I think doesn’t really look like good to me but please check it out and let me know uh what do you think uh about this I check Basque 2 which I think it is just maybe one typo away from being correct anyway so look uh not

09:53 a bad model at all for general purpose day-to-day task. I think it is uh quite a good daily driver. Let me know your thoughts in the comments. Again, if you want to help out, please become a member. Thank you for all the