New #1 open source AI model is here! FABLE LEVEL
Video Overview & Insights
Kimi K3 review. Best open source AI model. #ai #ainews #aitools #agi #llm #singularity Thanks to our sponsor Luma. Try Luma today: https://lumalabs.ai/aisearch-agents20
Kimi K3: https://www.kimi.com/
Kimi attention residual explainer: https://youtu.be/2IfAVV7ewO0
Its okay to steal from china, cause china steal from you - Pewds
Also see:
GLM 5.2 review https://youtu.be/6d__WOpZswY
damm
GPT 5.6 review https://youtu.be/SettwwX2cCI
Claude Fable review https://youtu.be/GUEE9OA4keo
Good Dario should never win
00:00 Kimi K3 intro
00:40 Kimi code
the v8 engine looks like shit lol what do u mean it looks beautiful
02:37 Liquid physics test
05:20 Blender v8 engine
Wow 😳
My goodness that is powerful
07:26 Financial explainer video
10:10 3D model test
Wow, almost all those things look like ass, will be interesting to see if anything actually good can be made in the future
12:25 Luma Agents
14:31 Mecha shooter
So, Gemini said that AI is Deus Ex Machina. Super AI's will eventually inhabit space, manage the Earth, and humans will become the "borg lite" by merging tech into their bodies. Brazil is the first nation on Earth to codify laws that protect a person's sense of "self" from being hacked aka "Brain Jacking" due to technical augmentation (putting microchips directly connected to the brain). The large datacenters being built that consume massive gigawatts of electricity are pivotal to the agenda.
I saved the transcript.
16:13 Music composition
19:11 Cancer detection
ЛОЖЬ.ИИ НЕ РЕШИЛА ЗАЛАЧУ 21 КАЧЕСТВ.ЗАДАЧА 21 КАЧЕСТВ.1-7 энерго центры 1-7 вибрационные тела 1-7 измерения человека и планеты. показать и доказать дополнения # ЧАСТОТЫ ДАННЫХ ТАБЛИЦЫ С.И СУХОНОС Таблица частот объектов Вселенной
## По данным С.И. Сухоноса с переводом размеров в частоты
Формула перевода: ν = c/λ, где c ≈ 3 × 10¹⁰ см/с
| № | Объект | Размер (см) | Частота (Гц) |
|---|--------|-------------|--------------|
| 1 | Максимоны | ~10⁻³³ | ~3 × 10⁴³ |
| 2 | Странные частицы | 10⁻¹⁵·⁵...10⁻¹³ | 3 × 10²³...9.5 × 10²⁴ |
| 3 | Нуклоны (протоны, нейтроны) | 10⁻¹³ | 3 × 10²³ |
| 4 | Ядра атомов | 10⁻¹³...10⁻¹² | 3 × 10²²...3 × 10²³ |
| 5 | Мезоатомы | 10⁻¹⁰ | 3 × 10²⁰ |
| 6 | Атом водорода | 10⁻⁸ | 3 × 10¹⁸ | Таблица 2. Спектр ЭМ частот 6-2-1 Уровень Название Диапазон частот (Гц) Период T (секунды) Пример длины волны 6Гамма~10¹⁹ – 10²²+~10⁻²⁰ – 10⁻²²< 0.01 нм 5Рентген (X-ray)~10¹⁶ – 10¹⁹~10⁻¹⁶ – 10⁻¹⁹0.01 – 10 нм 4УФ (UV)~10¹⁵ – 10¹⁶~10⁻¹⁵ – 10⁻¹⁶10 – 400 нм 3Солнечный спектр (видимый)~4×10¹⁴ – 7.5×10¹⁴~1.3×10⁻¹⁵ – 2.5×10⁻¹⁵400 – 700 нм 2Тепловой / ИК (эфир-тепло)~10¹² – 4×10¹⁴~10⁻¹² – 10⁻¹⁴700 нм – 1 мм 1Атом водорода (первый кирпич)характерные ~10¹⁵ (переходы) Атом водорода | 10⁻⁸ | 3 × 10¹⁸ |
20:43 Find the frog
21:42 Deep research
How can we run such AI locally without paying much cost. Is it even possible and what specs of local Machine would be required..
24:25 Specs and benchmarks
28:31 Open source
too bad you can’t run the open source model locally without a million dollars worth of hardware
Newsletter: https://aisearch.substack.com/
Find AI tools & jobs: https://ai-search.io/
It's not "better than Claude" for medical images. 85% wrong for medical advice is terrible, and Claude is better in this case, as it refuses to give you wrong information. See it as brain surgery. The better model is the one which will refuse to attempt surgery with 85% failure rate.
Support: https://ko-fi.com/aisearch
Here's my equipment, in case you're wondering:
how much k3 cost to use?
Copy
Share
Task paused due to system peak Continue Task ... lol
Lenovo Thinkbook: https://amzn.to/4jWeKwH
Dell Precision 5690: https://www.dell.com/en-us/dt/ai-technologies/index.htm?utm_source=AISearchTools&utm_medium=youtube&utm_campaign=precisionai#tab0=0
Kimi 3 is impressive, but it just quadrupled in price ($15 per million output token). While great model (especially coating in front end), it's still below Sol and Fable on a lot of tasks. Not to mention, slow. So at that price, I'm not quite sure it still has the same 'value' compared to some of the other open source models. Just look at the omniscience vs cost per task metric (rewards accuracy, punishes hallucinations, forgives abstention). Kimi 3 is actually one of the least cost effective models on that board.
GPU: Nvidia RTX 5000 Ada https://nvda.ws/3zfqGqS
Mic: Shure SM7B https://amzn.to/3DErjt1
AI 🫧🫧 about to burst 😂
Audio interface: Scarlett Solo https://amzn.to/3qELMeu
Why not generate a single high graphics game and let's see.
More User Perspectives
its asking for subscription
@editzewWhere is the frog, guys?
@severusfate600The gov cant be gateholders of AI... other countries will just override them.
@RighteousBruceI didn't pass the frog test too and i paused and really tried 😂
@D3r4g0723:50 was all the information in that paper correct? And if possible, please let us know how much your prompts cost (eg. what do you pay for 12 M tokens)
@mickmack8026Hard to justify as I keep getting "Task paused due to system peak. Continue ". Plus it's $19/month, around ($27.20 AUD). Think this AI is sleeping.
@theaussiegeekI am so glad open source EVEN EXISTS!
@tony_solarThank god for China. Burst the bubble please!
@devilmaycryhardMy concern is that it might be too resource-intensive and require some seriously powerful hardware; let's hope they implement some optimizations.
@TimelessX974the cost of generating a game is heading towards the cost of buying it, when that happens any game you want just name it :D
@l.halawaniI'm still using *GLM 5.2 Max**. I even bought **$1,000* in credits. 😅
Why?
* ~50% cheaper than Kimi K3.
* Lower hallucination rate (28% vs. 51%).
* The reasoning and understanding are genuinely impressive.
I have *ChatGPT Pro* and *Claude* subscriptions too, and I regularly test all the top models. Every month I pick the best price-to-performance model for my company's developers, and so far GLM has stayed on top.
My take:
* Google Gemini 3 Pro: 👎a poop don't even understand sometime and keep looping at a bad point i think it has 99% hallucination rate btw i used free plan for testing this one .
* GPT-5.6 Sol: Good.
* Claude Fable 5: Better, but way too expensive to run 24/7.
For everyday coding and agentic tasks, *GLM 5.2 Max* is still my go-to. Claude is reserved for special cases where the extra cost is worth it.
as a 3d modeler, you should have showed us how the typology looks like on that engine, holy crap
@calamitistwhere was the frog!
@AliAlBahadilyReverse engineered*. And fake.
@Und34d_Rai key is visible remove it .
@aman.40399I wouldn't be too happy yet, people are gaming the benchmarks now. I've tried deepseek for coding projects and got poor performance when they're benchmarks are "on par". I've tried kimi 2.5 and also thought it was significantly behind opus 4.8 as much as I shit on that model as well. I think a lot of benchmark cheating has been occurring across the board though this has been ancedotal. Though I like the current output from your examples.
@Ortianewait where is the frog? wtf you cant just make a video about a frog in an image that the AI cant find when a human cant even find it wtf? was it lower left where it looks like it as half cut and pasted into the image?
@95greenbugsick!!!! really like your test, especially the frog and the tumor test. I am switching to Kimi 3 now!
@jackzeinder1705Can use it in mobile
@ShadowMax-d6vBut its still quite expensive
@nukegamereactor7750Good that we have China in this world. 😂
@ay5523I’m too poor to install ts😂
@JtheSexyBoyYou're baby sitting it, like "use hyperframe" and then you give a repo, you shouldn't need to add anything else.
That's not artistic guidance, that's workflow guidance.
Sam Almant, Dario Amodei, Elon and Trump may be throwing a tantrum at this moment
@PitiplamThis blender thing is fire😮
@bhaveshahire3687imo the frog test wasn’t fair because fable knew it was there while Kimi didn’t. You should’ve tried it again but now with the same conditions as fable had. And you had it right, it is our (at least My) favourite test. I was wondering the whole time when that would come and my heart skipped a beat when you finally brought it up!
@innovativeyetaknowledgless9035ASK IT WHAT HAPPENED IN 1989 Tiananmen Square. LET'S SEE HOW OPEN IT REALLY IS.
@SayWhaaaaaaaaaaaaaaaaaaaaaaatSo 1 million token window, but 17 million tokens to create a project. That seems like it would forget a lot along the way. Maybe that is why these models create errors some times?
@KDawg5000If the model is downloadable and can be used locally, why the hell do you need to keep quoting the amount of tokens? Its a pointless exercise.
@PIERCED6966"I hope you can feel the intelligence." No. There is no intelligence in any AI model. This is next token word prediction and algebraic matrix math, not reasoning or thinking. It seems to work only because you suspend your disbelief.
@JimAmosallowing chinese ai do whatever it wants on my pc?? thank you..but no
@SayWhaaaaaaaaaaaaaaaaaaaaaaatyeha yeah china did it...right :D
@SayWhaaaaaaaaaaaaaaaaaaaaaaatW opensource.
@TërnocBesaIs open source means free or not
@AkshayPratapSingh-c1p