LLM Parameters: What 70B Means and Which Dials You Can Turn

Updated in September 2026 (Published on November 27, 2024)

The term LLM parameters means two different things. Depending on your problem, you are looking for a different answer.

On the one hand, it refers to the parameter count. This is the amount of learned data in a model. You recognize it by names like 7B or 70B. On the other hand, it refers to the setting parameters. You provide these values with every request. These include Temperature or Max Tokens.

The parameter count determines the capabilities of the model. The setting parameters control the behavior in your specific case. This article first explains the parameter count. After that, we show you each setting parameter individually.

What are LLMs anyway?

Large Language Models are large neural networks. They have learned from massive amounts of text. Their main task is predicting the next word. Exactly how this technology works is explained in our article on how large language models work.

Parameter count: what 7B, 70B and 1T mean

The parameter count shows how many details a model learned during training. It determines the knowledge of the AI.

Imagine parameters as small adjusting screws. The model adjusts these screws during training. A model named 7B has 7 billion such screws. 70B means 70 billion screws, and 1T stands for one trillion. These screws store the knowledge of the model.

Do not confuse this with training values like the learning rate. The developer sets these values during training. They are unimportant for you as a user.

What size means for quality, hardware and costs

More parameters usually mean more world knowledge. The model speaks more languages and plans more logically.

A small model is like an agile city car. It runs fast, costs little and fits on normal computers. A large model is like a heavy truck. It solves complex tasks better and speaks more languages. In return, it is slow and expensive to run.

You feel the size in everyday life in three places:

1. Price: Larger models cost more money per word.
2. Runtime: The model needs more calculation steps for each word.
3. Memory: An 8B model fits on a good graphics card, while a 70B model needs expensive special servers.

Details about hardware can be found in the article on local LLMs and open source models. Open models usually carry their size in their name. The large commercial providers often do not reveal the exact number.

Why setting parameters are worth it

Different settings give you completely different results for the same prompt. This makes the model flexible.

For an extraction from invoices, you want the same result every time. For an advertising text, you are looking for ten different suggestions. You use the same model for both. Anyone who only changes the prompt gives away half the possibilities.

All setting parameters at a glance

The table shows the most important adjusting screws. You see the effect and the names for the three major providers.

One development is important for your code. The classic dials lose their function in new models. Anthropic ignores Temperature and Top-P for models after Claude Opus 4.6. Google does the same for the Gemini 3 models. For OpenAI, these fields are also locked for reasoning models.

For older models and local systems, everything that follows applies unchanged. Which model knows which parameter is shown in our LLM capability matrix.

Controlling creativity

With these values, you determine how boldly the model selects words.

Temperature

Temperature changes the deviation from the most likely word choice. Low means predictable, high means surprising.

We test the prompt "Describe a sunny day in the park" with three values.

At Temperature 0.2: The sky is clear, and the sun shines brightly. It is quiet and peaceful. At Temperature 0.7: The sun shines warmly on the green meadows, and children laugh on the playground. At Temperature 1.0: The sun dances on the waves of the small pond, while colorful kites float in the sky.

With a value of 0.2, you get ten very similar answers. At 1.0, you receive ten completely different texts.

Top-P

Top-P limits the selection based on the accumulated probability. The model only chooses from the best words.

With a value of 0.7, the model collects the best words until they reach a combined 70 percent probability. This is flexible. If the model is certain, few words remain. If it is uncertain, it chooses from many words.

Top-K

Top-K cuts off the selection at a fixed number. The model ignores all other options. With a value of 10, the model only considers the ten best words. The exact probability does not matter. This helps against completely absurd terms.

Never change Temperature and Top-P at the same time. Both values trim the same selection. You will not know afterwards which value had an effect. Use Temperature as the main dial and leave Top-P at 1.0.

Controlling length and stopping

You determine with these settings when the model stops writing.

Max Tokens

Max Tokens is a hard upper limit for the answer. The model does not write shorter, but stops abruptly. A token is not a whole word here. German words often break down into several tokens. The token counter shows you the exact value for your text.

Every token costs money. Compare the costs for different lengths in the LLM price calculator.

Stop Sequences

Stop Sequences end the text at a specific word. The model stops writing immediately. For example, you set the word "User:" as a stop sign. As soon as the model generates this word, it stops. This helps enormously with lists or fixed templates.

Context window

The context window is the maximum space for the question and answer combined. It is not a dial, but a fixed limit. If the text is too long, the model stops. Sometimes it also forgets the beginning of the conversation. For long documents, calculate beforehand whether everything fits in.

Dampening repetitions

These penalties prevent the model from always using the same words.

Frequency Penalty

This value penalizes frequent repetitions. The more often a word occurs, the more unpopular it becomes. Without penalty: The cat is a cat that is often outside. With penalty: The cat is beautiful and loves to be outside.

Presence Penalty

This value takes effect at the second occurrence of a word. This forces the AI to use new terms. The model avoids a word as soon as it has been mentioned once. As a result, the text moves further away from the topic in terms of content.

Without penalty: The sunset is beautiful. The sunset is red. With penalty: The sunset is beautiful. The colors are impressive.

Repetition Penalty

This penalty is often found in local models. It pushes the probability for used words directly down. Values around 1.0 mean no penalty. Values above 1.2 often destroy fixed terms like product names. For models with strong repetitions, this value helps better than the other penalties.

Reasoning and thinking budgets

Current models think before answering. You control the duration of this consideration.

These models first generate invisible thinking blocks. After that, they write the visible answer. You choose a level or a fixed token count for this. The exact names are in the providers' manuals.

You need to know three important points:

1. Thinking costs money. A high budget increases the bill significantly.
2. Thinking takes time. The user waits noticeably longer for the answer.
3. Thinking and writing often share a common limit.

For Anthropic, thinking counts towards the Max Tokens limit. If you do not increase the limit, the answer remains empty. This happened to us in our own bulk run. A high budget pays off for mathematics or code. For simple tasks like formatting, thinking is skipped.

Forcing format: Structured Outputs, JSON mode, tool calls

You can make answers readable for machines. There are three different ways to do this. Structured Outputs force a fixed schema with exact field names. JSON mode only guarantees a valid format. Tool calls let the model execute a function.

For databases, you use Structured Outputs. JSON mode is a workaround for older models without a schema function.

Further API settings

These values control the system in the background. They do not directly change the language.

Seed only makes results similar, not identical. Prompt Caching only pays off for very long texts. Check the documentation to see if your model supports these values.

Presets for typical tasks

Use these proven starting values for your tests. Adjust them to your exact task afterwards. The table applies to models with classic dials. If these dials are missing from your model, just pay attention to the thinking budget.

Always collect twenty real examples. Then change exactly one value. Check the new results. A single attempt proves nothing at all.

What parameters do not solve

Parameters do not make a bad prompt good. They only reinforce your instructions.

A value of 0 does not make answers true. A model then simply invents facts the same way every time. Only your own data helps against false facts. You can read more about this in the article on AI hallucinations.

An unclear prompt remains unclear. If the target audience is missing, even long thinking will not help.

Here you will find quick answers to the most important questions.

Conclusion

Always proceed in the same order for your projects. This saves time and money.

First, choose the model. Then write a clear prompt. After that, set the thinking budget. Only change the Temperature at the very end. Note down values that work well. This way, you quickly build your own list.