They Trained an AI for Just $146 — I Tested It!

Author: Prompt Engineer 48 · Watch on YouTube

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: They Trained an AI for Just $146 — I Tested It! · Channel: Prompt Engineer 48 · 09:52 · Watch on YouTube

Topics: AI evaluation and benchmarks, AI models, Local and open-source models, Privacy and data protection

Summary

The video tests a small chat model that its developers say cost $146.26 to pre-train and can run locally in a browser or on a personal computer. The presenter explains its training, open release, and ways to try it, then evaluates it on an RTX 4060 using 13 prompts. The model is fast and gives some correct answers, but struggles with arithmetic, memory, Hindi, and following certain instructions. The presenter concludes that it is useful for learning, research, and offline experimentation, but is not a replacement for a larger chatbot.

Key points

The developers report that the main pre-training run cost $146.26, using eight A100 GPUs for 13.2 hours; the full project, including failed attempts, cost about $194.
The 152-million-parameter model is distributed as a 163 MB file and can run locally through a browser, Ollama, LM Studio, llama.cpp, or Apple’s MLX.
The developers published model repositories, data sets, training code, a technical report, and itemized evaluation answers.
On the presenter’s RTX 4060, Ollama reported speeds of about 300–426 tokens per second across 13 tests.
The model answered the capital-of-Australia question correctly and produced a correct prime-checking function, but added faulty code and an incorrect explanation.
It failed the arithmetic and bat-and-ball tests, could not recall details across turns, and did not respond appropriately to a Hindi prompt.
Thinking mode made the notebook arithmetic answer worse in the presenter’s test; results also varied with settings.
The presenter used a temperature of 0.3 and forced thinking off for most tests, and cautions that 13 prompts are only a small sample.
The developers report weak instruction-following results and disclose setbacks including a failed reinforcement-learning run, validation leakage, and deleted checkpoints.
The presenter recommends the model for learning, research, and offline building, but not for dependable arithmetic, long reasoning, or memory.

Chapters

Start	Chapter	Summary
00:00	The $146 model and its open release	The presenter introduces the small local chat model, its reported training cost, architecture, data, and the materials its developers made public.
03:09	Running the model and testing its answers	The presenter runs the model through Ollama and tests its speed, arithmetic, factual answers, code, language ability, memory, and instruction following.
07:19	Test limitations and reported results	The presenter explains the limits and settings of the 13-prompt test, reviews the developers’ reported evaluations, and describes disclosed training setbacks.
08:51	Who the model is for	The presenter weighs the model’s offline privacy and accessibility against its quality limitations, and suggests fine-tuning as a possible next step.

Quotes

$146. That is what the team says it cost to train this whole AI model from scratch — 00:00

It sounds confident and it is wrong. — 04:12

13 prompts is a tiny sample. — 07:19

I have rarely seen a release that is this open about breaking things. — 08:51

What you give up is quality, and now you know exactly how much because I showed you the wrong answers, too. — 09:22

Concepts and entities

Laghu, UGEAM Labs, Hugging Face, Ollama, LM Studio, llama.cpp, Gemma, Falcon, RunPod, RTX 4060

Full transcript (original language)

00:00 $146. That is what the team says it cost to train this whole AI model from scratch, not fine-tune it, train it from a random starting point. It is called LLaMA. It has 152 million parameters, and the file is 163 megabytes. It runs inside a browser tab. Now, here is the pain this is aimed at. Every time you ask a chatbot something, your words leave your laptop, they go to a server you do not control, and you pay for it in money or in privacy. And if you ever wanted to build your own model, the number people quote is millions of dollars. Most of us never even try.

00:31 LLaMA is one small team’s answer, a chat model you can run on your own machine with thinking you can switch on or off, and they published everything, the code, the checkpoints, and even their failures. But is it actually any good? I’m going to run it on my laptop, an RTX 4060 with 8 GB, and test it honestly, including the parts where it breaks. >> [music] >> Hey, I’m Prompt Engineer. Let’s find out what a $146 AI model can really do. First, who made it? The team is called UGEAM Labs, and LLaMA is the Sanskrit

01:03 word for small or light. The announcement went up on Hugging Face on October 22nd, 2026. Everything I quote from this page is what the developers report. I have not reproduced their training. What I can check is what runs on my own machine. In plain words, LLaMA is a small English chat model. It holds multi-turn conversations. It takes up to 16,000 tokens of input, and when you switch thinking on it writes out its reasoning before it answers. The recommended file is 163 megabytes. It also runs in Ollama, LM Studio, llama.cpp, and Apple’s MLX. Now, the number that

01:34 made me stop, the main pre-training run cost $146.26. That was eight A100 GPUs for 13.2 hours. The whole project, including the failed attempts, came to about $194. Compare that with the millions of dollars people usually talk about for big models. What did that money buy? The model read 52.6 billion tokens of text taken from 14 public sources whose licenses allow training. They trained their own tokenizer with 32,768 tokens, and before training, they removed every document that overlapped

02:05 with the test. They report 21,000 documents, so the scores stay clean. There are three stages. First, pre-training, where it learns to predict the next piece of text. Then, a long context stretch, so it can handle long inputs. And last, chat tuning on 41,467 example conversations, which is what turns it from a text continuer into something that answers you. The shape is borrowed on purpose. Lagu uses Google’s Gemma for layout, scaled down, so the tools that already run Gemma for run Lagu with no custom code. It has 28 layers in three of every four layers. It

02:37 only looks back over the last 512 tokens, and that keeps memory tiny. They report that the whole server needs about 154 MB at 4,000 context and 264 at 16,000. And then, what is open? Seven model repositories, two data sets, a live space, the training code on GitHub, a technical report, and every evaluation answer item by item. You can read the exact conversations it was tuned on for 146 dollar model. That is a real amount of transparency. If you want to try it yourself, there are four easy paths. The browser page, udgmlabs.com/models/lagu,

03:09 nothing to install, the model downloads once and runs on your device. The Hugging Face page, which runs on a free CPU and has a think first tick box. Ollama, one line and you add {dash} {dash} think equals false for normal chat. And llama.cpp, where you must keep {dash} {dash} ginger because without it, the server fails. Okay, enough reading. Let’s run it. One line in Ollama, Ollama run hf.co/udgmlabs/lagu-gguf:q8_0. It pulled 162 MB. For a language model, that is almost nothing. With it loaded,

03:40 my GPU showed 461 MB in use in total. First, a simple hello. I asked, “Hi, what can you help me with?” It says, “I’m here to help with a wide range of tasks.” And then lists problem solving and writing again and again six times with slightly different words. On topic, polite, a bit repetitive. Then I asked, “Who are you and who made you?” It said, “I’m Lago, a small language model built by Uqbar Labs.” Correct. Now, speed. Ollama reports its own tokens per second. Across my 13 tests, it landed between about 300 and 426 tokens per second on the 460. Answers appear almost

04:12 instantly. This is the whole point of a model this small. Now, the first real test, a simple one. What is 84 * 3 / 2? The correct answer is 126. Lago answered 66. No steps, no hedge, just the wrong number. So, here is the first honest result. It sounds confident and it is wrong. Next, the classic bat and ball puzzle. A bat and a ball cost $1.10. The bat cost $1 more than the ball. How much is the ball? The right answer is 5 cents. Lago sets up an equation with the bat as 1.10 * B and ends with the ball is 0.10 B.

04:45 That is not an answer. Wrong. With Ollama’s default settings, it did say 5 cents and I’ll come back to that. Now, the thinking switch. Same question both times. A notebook cost $3. How much do seven cost? With thinking off, it says 7 * 3 is $21. Correct. With thinking on, it reasons, multiplies three by 0.07, and answers 3 cents. Wrong. The article itself warns that thinking only helps on short arithmetic and that it hurts conversation. So, leave it off on my test. Thinking on made it worse. Then, facts. I asked for an interesting

05:16 fact about the universe. It did not give me one. It replied, “Sure. Which topic are you interested in?” And offered four options: black holes, galaxies, stars, and the cosmic microwave background. Not wrong, but I asked for a fact and got a menu. A short factual question. What is the capital of Australia? It said Canberra. Then it added that Sydney is the largest city and Melbourne is the second largest. All of that is correct. A clean pass. Code. I asked for a Python function that checks if a number is prime. The function it wrote is correct. It loops

05:46 up to the square root and checks divisibility. Then it added a line that would crash because the variable n is not defined there and an explanation saying the square root gives the largest even number, which is nonsense. So, it can produce a famous snippet, but it cannot reliably explain it. Languages. The word laghu is Sanskrit. So, I ask in Hindi written in English letters for a short story. It did not write a story. It said “Mujhe is the name of the person and choti is the name of the dish.” Mujhe means to me and choti means small. The model is

06:17 described as English only, so this is not a surprise, but it is worth knowing before you promise anyone multilingual. Multi-turn memory. I said, “My name is prompt and my favorite color is green. Remember that.” Then I asked, “What is my name and my favorite color?” It could not answer. Both replies were about the name prompt being a pun and suggested names like pretty. It did not remember the color at all. The developers report this category is only level with the similar small model and in my small test it failed. And an instruction, “Give me exactly three bullet points on why sleep

06:48 matters. Nothing else.” It gave exactly three bullets and the content is sensible, but then it added a sentence after them saying they do not add up to three bullet points. So, it broke the nothing else rule. This matches what they say, “Instruction following is weak.” Let me put my own scoreboard together. 13 tests. Clean passes on who are you, the capital and the notebook with thinking of mixed results on hello, which repeats itself, the three bullets with one extra sentence, the code, the universe fact request and quantum computing. And real failures on the arithmetic, the bat and ball thinking on

07:19 multi-turn memory and Hindi. This is a small sample, not a benchmark, but it matches the shape of what the developers report. A few limits on my own testing. 13 prompts is a tiny sample. I used a temperature of 0.3, not the default of 0.7, so answers vary from run to run. Lama turns thinking on by default, so I forced it off as the article recommends. When I first ran the same prompts with the default, the bat and ball came out, right? 5 cents though, the reasoning was garbled and the capital answer added a wrong claim. Same model, different settings, different results. I also did

07:49 not verify their training cost or their benchmark scores. I only checked that it downloads, runs, and what it says on my machine. Here’s what they report. Again, small LM 2, 135 million parameter instruct model, Laghu wins 51% of judged chat comparisons. That is level on grade school maths with thinking 7.9% against 1.8, but their own instruction following score is 26.1, which they call weak, and maths at 7.9% is low in absolute terms. And this is the part I respect most. Before training, they wrote down the bar a result had to clear. Laghu cleared

08:20 zero of three categories. A model from Falcon, 90 million parameters, trained with about 12 times the compute and a preference tuning stage is ahead on all three chat categories. They published Laghu anyway with no best-in-class claim. And they list what went wrong. A reinforcement learning run failed. A validation split leaked about 37% of validation conversations had a near copy in training. A storage cleanup on October 21st deleted most of the pre-training checkpoints, and they say that was their error. The main run even died at step zero, which cost about $2.

08:51 I have rarely seen a release that is this open about breaking things. So, who is this for? Not for replacing a big chatbot based on my tests. Do not trust it for arithmetic, long reasoning, or memory. It is for people who want to learn how a model is actually built because every checkpoint and recipe is open. It is for researchers who want a tiny to fine-tune, and it is for builders who need something that runs offline in a browser at 300 tokens a second, and where a wrong answer is cheap. Go back to the pain point. The price of entry to training your own

09:22 model just dropped to the cost of a nice dinner for four. The privacy cost of a tiny local assistant is zero because nothing leaves your device. What you give up is quality, and now you know exactly how much because I showed you the wrong answers, too. What I would try next is a fine-tune because the base models and the tuning data are all there and the space on hugging face runs it on a free CPU if you want to pocket yourself. Tell me in the comments, what is the one question you would ask a 150 million parameter model? If you’re looking for cheap and reliable

09:52 GPUs, I use RunPod and the link is in the description. If this was useful, the Ollama command and every page I showed are in the description. Thanks for watching and I will see you in the next one.