Parameters — 70B, 405B… What Do These Numbers Mean? (A Complete Guide to AI Model Size and How to Choose a Model in 2026)

The B after a model's name is not a report card; it is closer to a hardware requirements sheet. The moment you read this number as a performance ranking, your model choice is almost guaranteed to be wrong.

Background / Why This Topic Now

Look at AI model names and the numbers after them catch your eye: 8B, 70B, 405B, even cryptic notation like 671B-A37B. Here B stands for billion, that is, 1,000,000,000. 70B means 70 billion, and 405B means 405 billion.

The problem comes next. Most people read this number like engine displacement or CPU clock speed, as a "bigger is better" performance metric . But in 2026 this reading fails for three reasons. Architectures that perform well without using many parameters have appeared, major commercial models do not disclose their size at all, and the same size yields completely different results depending on how it was trained.

If Part 1 of this series was about AI being wrong , Part 2 about it being inconsistent , and Part 3 about what it can take as input, this installment is about what to look at when choosing an AI.

Key Data & Current State

The reality of size: Llama 3.1 405B has 405 billion parameters, was trained on more than 15 trillion tokens, and needs 810GB at FP16 just to load its weights into memory(203GB when compressed to INT4) (Meta / Hugging Face, 2024)
The study that overturned the size race: GPT-3 was trained on 300 billion tokens with 175 billion parameters, 1.7 tokens per parameter, whereas DeepMind's Chinchilla trained 70 billion parameters on 1.4 trillion tokens, 20 tokens per parameter, and performed better (Hoffmann et al., 2022)
Total vs. actually used: DeepSeek V3 has 671 billion parameters in total, but what actually activates to process a single token is about 37 billion (5.5% of the total) (per the model card and architecture explainers)
A number that cannot be known at all: In the GPT-4 technical report, OpenAI stated that "this report contains no further details about the architecture (including model size), hardware, training compute, dataset construction, training method" (OpenAI, 2023)

? Interpretation: The second line is where this article starts. The smaller model got the better score. And it was no fluke or exception; it was a systematic conclusion that, for the same compute budget, increasing data beats enlarging the model. Parameter count is just one of several variables that determine performance, and it is meaningful only when paired with data and training volume.

In-Depth Analysis

1) What Exactly Is a Parameter?

A parameter is each individual numberthe model adjusts on its own during training. They are often compared to synapses in the brain, but in practical terms it is more accurate to think of them as adjustment knobs. Training is the process of turning these hundreds of billions of knobs little by little to find "the setting that best predicts the next word."

So what the parameter count signifies is not the size of intelligence but the total amount of patterns the model can hold, in other words, capacity. A bigger bowl does not make the food taste better, but a bowl that is too small means some dishes will not fit.

How to feel the number physically: One parameter usually takes 2 bytes (FP16). Then the math becomes very simple.

8B model → about 16GB
70B model → about 140GB
405B model → about 810GB

Recall that a high-end consumer graphics card has around 24GB of memory, and what 405B actually means hits home immediately.
It is not the size of the performance. It is the size of the bill.

2) Why Was It Once a Numbers Race — and What Overturned It?

In the early 2020s the industry was dominated by the observation that making models bigger improved performance predictably (scaling laws). So the parameter count was the headline of every announcement.

What put the brakes on this trend was DeepMind's 2022 Chinchilla study. The researchers recalculated the balance between model size and training data volumewithin the same compute budget and concluded that the large models of the time were seriously undertrained. By their calculation, for GPT-3 to be optimally trained it needed not 300 billion tokens but about 3.5 trillion (Hoffmann et al., 2022).

Author's view: The real meaning of this study is not "small models win." It is that comparing parameter counts in isolation became meaningless. The same 70B becomes a completely different thing depending on how much data it was trained on and how well. Same displacement, different engine generation, different car.

3) Three Reasons You Can't Trust This Number Today

① MoE — total and actually-used have split apart

Recent large models widely use the Mixture of Experts (MoE) architecture. Several "expert" blocks sit inside the model, and for each input token a router activates only some of them. The rest sit in memory but take no part in the computation.

This is why notation like 671B-A37B appears. The first number is total parameters; the number after A is active parameters per token. In DeepSeek V3's case, only about 37 billion of the 671 billion total actually operate. Calling this model "671B-class performance" and calling it a "37B-class model" are both inaccurate. Memory follows the total; speed and compute cost follow the active count.

② Undisclosed — there is no number to begin with

In the GPT-4 technical report, OpenAI stated it would not disclose details including model size, citing the competitive landscape and safety (OpenAI, 2023), and the major commercial models since have largely kept the same stance. The figures circulating online, "so-and-so has X trillion parameters," are mostly estimates or unsourced rumors. In other words, the models we use every day are precisely the ones whose number we cannot know.

③ Post-training — same size, different results

Fine-tuning after pretraining, alignment training, and how long the model is allowed to "think" before answering all change the results of a same-sized model dramatically. Parameter count is only the starting point of all these processes, not the destination.

4) So When Is This Number Useful?

This does not mean parameter count is meaningless. It just serves a different purpose.

What you want to know	Can parameter count tell you?
How smart is this model?	❌ (check with benchmarks and direct testing)
Will it run on my hardware?	⭕ (total parameters × precision)
How fast will it respond?	⭕ (for MoE, based on active parameters)
Why is the API priced this way?	⭕ (roughly proportional to size)
How much will fine-tuning cost?	⭕

In short, parameters are an infrastructure metric, not a performance metric. The number tells you not "how well it does" but "how much it costs."

In Practice — 6 Principles for Reading the Number Properly

① Check memory with mental math

Required memory (GB) ≈ parameters (B) × 2 (at FP16). Quantize to 4-bit and it shrinks to roughly × 0.5. On top of that you need cache memory to hold the conversation context, so in practice leave headroom above the calculated value.

8B model → about 16GB at FP16 / 4–6GB at 4-bit (feasible on a typical laptop or consumer GPU)
70B model → around 40GB even at 4-bit (workstation class)
405B model → over 200GB even at 4-bit (multi-GPU server)

② Look for the number after the "A" first

When looking at MoE models, active parameters are closer to the real-world feel than total parameters. Size memory by the total; estimate speed and cost by the active count , and you will mostly be right.

③ Don't compare cloud models by parameters

Service-based models like ChatGPT, Claude, and Gemini do not disclose their size, so they are not comparable on this axis to begin with. In their case, benchmark scores, price per token, response latency, and results from running your own work through themare the only meaningful metrics.

④ Choose by task difficulty

Not every job needs the largest model. On the contrary, most practical work is well served by small models.

Small (1B–14B) is enough: classification, tagging, information extraction, structured summarization, format conversion, typo correction
Large has the edge: multi-step reasoning, complex coding, synthesizing long documents, interpreting ambiguous instructions

⑤ Think of cost as "size × number of calls"

Calling a small model several times is often cheaper and faster than calling a large model once. Conversely, if the small model fails and you end up rerunning on the large one, you pay twice. A tiered setup that tries small first and escalates to large on failureis the safest bet in practice.

⑥ In the end, verify with your own data

The principle from Part 3 applies here unchanged. Public benchmarks are a reference; the basis for judgment is the results of running 20–30 samples from your own work. A model chosen by testing on real documents is almost always better than one chosen by parameter count.

Implications

EXAONE 4.5, released by LG AI Research in April 2026, has 33 billion parameters, about one-seventh the size of the company's largest model, yet by LG's own account delivers performance aimed at interpreting complex documents such as contracts, technical drawings, and financial statements (LG AI Research, 2026). Instead of growing the size, the direction is narrowing the purpose.

There are reasons this approach fits the Korean environment especially well. First, small and mid-sized models released as open weights open up the option of processing documents that cannot leave the company in-house. Second, in narrow and repetitive domains like Korean business documents, specialization in that format may translate into results more directly than the broad knowledge of a general-purpose giant model.

Author's view: The most common mistake Korean companies make when evaluating model adoption is the judgment that "using the largest model is the safe choice." The real bottleneck is usually not model size but the state of the company's own data. For the same budget, a mid-sized model plus well-organized internal documents often produces better results than a larger model.

Outlook & Variables to Watch

Improvement variable — deepening sparsity: As designs that "build big but use little," like MoE, spread, the gap between total parameters and actual compute is likely to widen further. In that case, listing total and active counts side by side will likely become the de facto standard.
Risk variable — incomparability: If commercial models keep their sizes undisclosed, benchmarks are the only basis on which consumers can compare models objectively, and benchmarks are not free of contamination and overfitting controversies. "The absence of a trustworthy basis for comparison" is itself the risk going forward.
Checkpoints: ① Whether disclosing active parameters becomes standard practice for major open models ② Changes in the upper limit of model size runnable on consumer hardware (unified-memory devices) ③ The size range of models chosen in on-premises adoption by Korean public institutions and companies

Conclusion

B is a spec, not performance. Parameter count tells you not "how smart" but "how much hardware and money it needs." Required memory is mental math: parameters × 2GB.
Size alone does not make a comparison. Data volume (Chinchilla), active parameters (MoE), and post-training all change the result. And the commercial models we use most do not disclose their size in the first place.
The criterion for choosing is the task, not the size. For classification, extraction, and summarization, start small and step up to a large model only when it fails. And the final judgment should always be the result of running your own work samples yourself.

Parameters — 70B, 405B… What Do These Numbers Mean?

Parameters — 70B, 405B… What Do These Numbers Mean? (A Complete Guide to AI Model Size and How to Choose a Model in 2026)

Background / Why This Topic Now

Key Data & Current State