A Guide to ELM’s Generative AI Models: How to Choose the Right One for Your Tasks
Quick recommendations
Llama 3.3
Choose if:
- you want the most sustainable option
- your task is straightforward
- you want a fast response
- you are doing everyday writing or support tasks
Perfect for simple office tasks. Switch if you need more capability for complex tasks, you work with longer documents or your need a model with reasoning enabled.
GPT-5.4-mini with medium reasoning
Choose if:
- you need a good balance of quality and speed
- you are working with longer documents or complex spreadsheets
- your task is moderately complex
To learn more about the Generative AI models available in ELM and ELM’s advanced capabilities, consider signing up for a training session via this link https://elxw.fa.em3.oraclecloud.com/fscmUI/redwood/learner/learn/catalog-details?launchedFrom=wtl&learningItemId=300002271837901&learningItemType=ORA_COURSE&tab=explore
Introduction
One of the most asked questions by users is which Generative AI model they should choose in Settings.
By default, you are assigned Llama 3.3 model. In ELM, this model is marked with a green leaf, indicating that it is the more environmentally friendly option among the models available through our partners.
For most users, Llama 3.3 will be sufficient for the majority of everyday tasks. However, ELM also provides access to more powerful external models for users working on more complex tasks.
The best model for you will usually depend on three things:
- how complex your task is
- how quickly you need your responses
- how much context the model needs to handle, especially when working with long documents
Sustainability & environmental impact
One of ELM’s core principles is to be climate-sensitive. ELM helps reduce the environmental impact of Generative AI by providing access to locally hosted, greener, open-source optimised Generative AI models.
We are committed to being transparent about our environmental footprint and are actively working to improve how we evidence our impact and mitigation strategies.
Why Llama 3.3 is the greener option
The Llama 3.3 model:
- does not rely on any external infrastructure
- does not use any water for cooling
- is powered entirely by renewable energy.
Approximately 60 queries to ELM using the Llama 3.3 model consume about the same amount of energy as one full phone charge.
For that reason, we strongly encourage users to use Llama 3.3 unless they have a clear reason to switch.
When should I use an external model instead of Llama 3.3?
We strongly believe that the majority of users can do very well with using Llama 3.3 within ELM, and that they do not need to switch to using the external models. It is also important to note, that Llama 3.3 requires more descriptive prompts to complete more complicated tasks, however, given that consideration it can often match the performance of external models.
For more advanced use cases, such as complex document analysis, programming, or working across very large files, a more powerful external model may be useful.
However, it is worth keeping in mind that these models:
- usually require more computation
- typically have a greater environmental impact
- are more expensive to run
- may be unnecessary for simple tasks
In particular, choosing the most powerful model with the highest reasoning setting enabled can significantly increase cost, energy use and response time. Where possible, choose the model and reasoning level that best match the task.
In most cases, there is little value in using the newest flagship model with the highest reasoning setting for every prompt.
Reasoning
Understanding your options
Once you select a model different to Llama 3.3 under settings, you might be presented with the option to select a reasoning level. This is because some of the models of our partners allow for this to be enabled.
What is reasoning in generative AI?
“Reasoning” in Generative AI means that unlike non-reasoning AI systems that produce a single, immediate response to a prompt, reasoning models allocate additional computational effort during inference to decompose problems into smaller steps and evaluate multiple possible solutions before attempting to select the most appropriate answer. Reasoning in Generative AI does not mean that suddenly the predictive models start “thinking” in the human sense of the word. It also does not guarantee that the answer will be correct, although at large it does improve the quality of the answers
Why not always choose the highest reasoning setting?
The simple answer is that more reasoning takes more time and more compute.
For straightforward tasks, high reasoning settings are often unnecessary. They increase waiting time and cost without producing a noticeably better result. For more complex or higher-stakes tasks, the extra time may be worth it. The right balance depends on what you are trying to do.
Illustrative example: reasoning, speed and cost
The table below shows how increasing reasoning may affect processing time and cost for the same input. It is also important to note that higher cost means more energy was used to generate a response
Note: These are illustrative estimates, not exact measurements. Actual cost depends on the model, provider pricing, input length, and output length. Because pricing can change, the table uses relative cost rather than fixed currency values.
| Reasoning amount | Example input size | What the model does | Example time | Indicative cost |
| None | 100 pages of text | Produces a direct answer with no extra internal steps | ~30 seconds | Lowest/ baseline |
| Minimal/Low | 100 pages of text | Performs a small amount of extra step-by-step evaluation before answering | ~1 minute | Low |
| Medium | 100 pages of text | Breaks the task into more sub-steps and considers more candidate answers | ~2 minutes | Moderate |
| High | 100 pages of text | Uses substantially more inference-time computation to explore and refine answers | ~3 minutes | High |
| Very high | 100 pages of text | Maximises internal problem decomposition and answer checking, prioritising quality over speed | ~5+ minutes | Very high |
If a task is time-sensitive or relatively simple, a lower reasoning setting is often the better choice. If the task is complex, high-stakes, or requires more careful analysis, a higher reasoning setting may be worthwhile.
Which Generative AI model should I choose in ELM?
Historically, some models were seen as better for STEM-related tasks and others as stronger for humanities-based work. Currently, it is not the case, however GPT models have an advantage in STEM related tasks over our locally hosted Llama 3.3 model, due to how they overall are stronger.
Overall, all models available within ELM are good for all your tasks, such as
- refining and drafting emails
- summarising and querying documents
- querying Excel spreadsheets (for example, asking for formulas or to write reports on data)
- brainstorming
- coding and scripting
Your only consideration when choosing a model for a specific task is complex the task is, how quick you will need answers and context size.
What is context?
In ELM, context refers to the amount of information a model can retain within a particular chat.
If you continue an existing conversation, that context is retained. If you start a new chat, the context resets.
For most users, context size will not be a problem in normal chat use. It matters most when you are:
- uploading long documents
- working across multiple large files
- asking the model to analyse long or complicated material in one conversation
If you exceed a model’s context limit, ELM will notify you in the interface. In that case, you may need to switch to a model with a larger context window.
Context is most important when uploading documents to ELM via the document to upload feature. Since most documents our users upload are bulky, it is possible to exceed the limit. If you do, you will be notified of this in our UI.
You do not need to understand the technical detail of tokens to make a good choice here. In practical terms, if ELM tells you that the context limit has been reached, choose a model with a higher context size.
Models most users are likely to need
There are several models available in ELM, but the following are the ones most users are most likely to use regularly.
| Model name | Context size | Speed | Capabilities |
| GPT-5.4 | 1,050,000 | Slower | Complex tasks, analysing very long and intricate documents and excel spreadsheets, programming. Overkill for simple office tasks, such as rewriting emails. |
| GPT-5.4-mini | 400,000 | Fast | Strong middle ground. With extra complicated tasks, you trade off some capability for speed. Quality drop is not easily noticeable for most complex tasks. |
| GPT-5.4-nano | 400,000 | Instant | Fastest answers, but not the best for complex issues. Good for more basic tasks. Less suitable for difficult analytical tasks, where the difference between it and the flagship model becomes more noticeable. |
| Llama 3.3 | 128,000 | Instant | This is the default model and the most environmentally friendly option. It is fast, reliable for common day-to-day tasks, and will be entirely sufficient for most users. Provides fastest answers. |
There are more models available within ELM, but realistically these are the ones you need to know. The other older models are available to serve our long-term users who got used to them. Eventually, access to older models will be reduced.
Recommended model and reasoning configurations
If you are set on using models with reasoning capabilities instead of the locally hosted Llama 3.3 model, here are configurations that will give you the best outcomes for most use cases.
| Model name | Reasoning amount | Speed | Use case |
| GPT-5.4-mini | Medium | Medium Slow | If you spend most time analysing long documents. |
| GPT-5.4-nano | Low | Extra Fast | If you spend most time supporting basic office tasks, such as drafting emails, adjusting tone of responses, brainstorming. |
| GPT-5.4 | Extra High | Slow | For most complex tasks, including long document and spreadsheet analysis. |
Conclusion
In conclusion, choosing the right Generative AI model in ELM depends on the complexity of your task, the speed of response you need, and the context size required.
For most everyday tasks, a default model will suffice, but for more complex tasks or larger documents, a more powerful model may be necessary. Considering the environmental impact and cost of using external models, it’s essential to strike a balance between model capability and resource usage.
By understanding your options and selecting the most suitable model and reasoning level, you can optimise your use of ELM and achieve the best outcomes for your tasks.
