Second Place is Desperate: A Look at China's AI Crisis in Beijing
- Haebom
2
Haebom's Archive
Coaching resumes.
- Haebom
🎄To celebrate Christmas, we are releasing articles previously available only to members for free.
- Haebom
🎄 People who have nothing to do on Christmas, come here
- Haebom
Curiosity
- Haebom
5
🟣 When the conversation starts again
- Haebom
🎖 Time Magazine's Greatest Inventions of 2025
- Haebom
It's the "hardware", stupid
- Haebom
1
Cost Golden Cross in the Content Market
- Haebom
Light things I made during the holidays
- Haebom
In SaaS, competition is about "systems," not "wills."
- Haebom
10 Dangerous Misconceptions About Vibe Coding: The Characteristics of Language Models
- Haebom
The KakaoTalk reorganization shouldn't be used as a trigger for a firestorm like this.
- Haebom
5
Marriage with an AI lover?
- Haebom
Vibe Coding Cleanup: The Birth of a New Tech Side Hustle Ecosystem
- Haebom
The Evolution of AI UX: From Invisible Magic to Inevitable Experiences
- Haebom
Things I thought about over the weekend
- Haebom
2
Albania Appoints World's First AI Minister: A Dilemma: The Reality vs. the Fantasy of a Corruption-Free Government
- Haebom
Meta Ray-Ban Display: Zuckerberg's Vision for the Future of Computing
- Haebom
Publication Announcement for "People Who Entrust Their Thoughts"
- Haebom
1
haebom
해봄의 아카이브
Why did Samsung choose Llama 4?
Haebom
Recently, I saw an article that Samsung Electronics Semiconductor Division (DS) has introduced Meta's latest large-scale language model (LLM) , LLaMA4 , to the entire organization. Personally, why did they choose LLaMA4, which is surrounded by performance controversies and even suspicions of benchmark manipulation ?
Samsung Electronics DS Selection and Its Background
The LLaMA4 introduced by Samsung Electronics is a multimodal AI consisting of Scout (lightweight model) and Maverick (basic model) . It is known as an excellent model that can process not only documents but also images, voices, and videos at once. In addition, it provides an overwhelming context length of 10M tokens, so it seems to be very attractive to Samsung, which handles large-scale data such as semiconductor process data. However, it is questionable whether this choice was actually the best. If we look at some realistic reasons, there are some regrettable points.
Performance and Benchmark Reliability Controversy
LLaMA4 has been the subject of controversy over its performance since its release. Some researchers and the developer community have raised suspicions that the benchmark model submitted by LLaMA4 differs from the actual published weights. There was even a revelation from an internal employee that "there was excessive adjustment to increase the benchmark score."
This controversy is a serious issue that companies must consider when handling important business data. For a company like Samsung, where process optimization and reliability are essential, it is hard to shake the feeling that this was a rather risky decision.
In fact, I personally think that the introduction of LLaMA4 was simply an option for employees, but the media is exaggerating the issue.
More efficient alternatives: the existence of Gemma3 QAT
In fact, there were alternatives on the market that showed excellent performance and efficiency in addition to LLaMA4. For example, Gemma3 QAT (27B) is evaluated as a model that shows excellent performance even with relatively small hardware resources. Of course, considering that QAT itself is an LLM that can run smoothly even with general consumer-level computing power... It seems a shame to pay for Maverick. A company as big as Samsung would probably use a paid model rather than a community version. (After all... Mr. Don)
Item | Llama 4 Scout (109 B/17 B Active) | Gemma 3 27B QAT |
Multimodal | Simultaneous processing of text, images, voice, and video | Text/Image (Voice X) |
Context length | 10 M tokens (Maverick 1 M) | 128K tokens |
Inference speed | 120 tps (FP8, H100 1 sheet) | 20-25 tps (INT4, RTX A5000) |
Memory | Active 17 B → FP8 24-32 GB | INT4 12-24 GB |
MMLU-Pro | 69.6% (Scout) 79.4% (Maverick) | 66-67% (27B IT) |
License | Meta Community (Paid for companies with MAU 700 million↑) | Apache-2.0 (no derivative works disclosure obligation) |
GPU scale | H100 1 sheet (Scout) / 8 sheets↑ (Maverick) | 3090 1 sheet possible |
•
Gemma3 QAT is a memory-efficient INT4 quantization method that can achieve sufficient performance with just one GPU.
•
Even in the MMLU benchmark performance , it only shows a slight difference of about 2-3%p compared to LLaMA4 Scout, making it a much more economical and efficient choice.
In the end, it's because of 'speed'?
The most realistic reason Samsung adopted LLaMA4 was likely because of 'speed' . If this was spread by the media, it was probably because they wanted to show that they are also doing something with AI, and if the media misunderstood it, it was just a decision made because they have a lot of money. What can you say when a company makes an expensive decision with company money? (Shareholders don't know.)
•
With competitors like SK Hynix and TSMC rapidly achieving AI-based process innovation, Samsung would not have had time to wait for its own model to improve. In addition, Hynix has already introduced AI into practice in the manufacturing sector, along with Gauss Labs...
•
Thanks to the MoE structure (active parameter 17B) of LLaMA4 , GPU efficiency was also secured, so fast task application was possible. In addition, seeing that the article says that there is a report that LLaMA4 showed quite good results in multi-modal tasks in Samsung's internal pilot test, I wonder if there is something there.
•
Personally, I think it may be because I have a preconceived notion about Maverick's performance and its lack of multi-modal audio and video processing capabilities.
What changes will Samsung's choices bring in the future?
It is not yet known how these decisions will be evaluated in the long term. Of course, productivity and process efficiency may increase immediately. However, internal verification and objective evaluation of models that have been introduced with issues such as benchmark reliability or performance exaggeration have become more important. However, I am not sure how to apply them to work unless they are used in a process-related or predictive manner like Gauss Labs. For now, the only way to use LLaMA4 is to create a similar ChatGPT or to create and modify documents...
Personally, I was curious about the background behind the decision and the thinking behind it, so I wrote it down. To be honest, I wanted to write something because GPT drew the illustration so well.
Subscribe to 'haebom'
📚 Welcome to Haebom's archives.
---
I post articles related to IT 💻, economy 💰, and humanities 🎭.
If you are curious about my thoughts, perspectives or interests, please subscribe.
haebom@kakao.com
---
I post articles related to IT 💻, economy 💰, and humanities 🎭.
If you are curious about my thoughts, perspectives or interests, please subscribe.
haebom@kakao.com
16
S
Snowy white dream
/
S
Snowy white dream
/
S
Snowy white dream
S
Snowy white dream
S
Snowy white dream
S
Snowy white dream
/
/
S
Snowy white dream
/
S
Snowy white dream
S
Snowy white dream
S
Snowy white dream
S
Snowy white dream
S
Snowy white dream
S
Snowy white dream
S
Snowy white dream
S
Snowy white dream
S
Snowy white dream
See latest comments