AIs

All Providers

Telosnex integrates with leading AI providers.

You can choose which AI providers to use in Settings > AIs.

The sections below describe each category of AI, and each provider of it, in detail, including cost.

LLM

Search

TTS

STT

Local LLM

Art

Pricing

Where possible, Telosnex offers free on-device AI options alongside cloud-based paid AI options. Web Search and Art are always cloud-based.

With a membership, paid AIs are provided at cost.

Without a membership, paid AI costs 4x as much.

Advanced users, such as developers, can provide API keys that instead use the billing they have with the AI provider.

There are 5 AI categories:

LLMs

LLM stands for Large Language Model, a particular type of AI like ChatGPT. They generate text. Some, known as multimodal LLMs, can also take in images and generate text.

Telosnex has options for both cloud-based and on-device LLMs.

On-device LLMs run for free on your device, but tend to be less powerful than cloud-based LLMs. They also require a significant amount of memory and CPU.

Cloud LLMs

This screenshot shows only 3 of 5 models

Cloud LLMs are the best choice, almost all the time.

The best quality AIs tend to be cloud-based. Local AIs require a powerful device, they push the device to its limits.

Telosnex offers 5 cloud-based LLMs:

Small $1 = 22 hours; Large $1 = 23 minutes
Anthropic's Claude 3 Haiku and Opus were released in March 2024.

Anthropic may be the best combination of quality and pricing. Groq, while cheaper and faster, runs Llama 3 which is of less quality. It struggles to produce structured responses, and thus generating UI, like ideas on the home screen. OpenAI has a worse quality and more expensive small model, GPT 3.5, the model behind ChatGPT.

Small $1 = 17 hours; Large $1 = 2 hours Gemini 1.5 was released in May 30th 2024. It is a midpoint in quality and pricing. It struggles to produce structured responses, and thus generating UI. It cannot generate ideas on the home screen.

Small $1 = 200 hours; Large $1 = 16 hours
Groq is new AI company that it runs other companies models on its specialized hardware. It is significantly faster (answers appear almost instantly) and significantly cheaper (10x compared to the next cheapest).

The downsides are, its best models, Llama 3, are somewhat lower quality than the best. It struggles to produce structured responses, and thus generating UI, like ideas on the home screen. It also has capacity issues, so it may not be available. (though, it is simple to temporarily switch AIs)

Small $1 = 11 hours; Large $1 = 1 hour
Mistral Small and Large were released in March 2024.

Mistral has won praise for its quality and ability to keep up with much larger models and companies. Mistral shares their AI models, making them popular with AI hobbyists as well as the enterprise. Their models are particularly good in multilingual conversations, a focus of Mistral, based in France.

Small $1 = 14 hours; Large $1 = 1.5 hours
The classic, industry-leading AI. OpenAI's GPT 3.5 is the model behind ChatGPT. OpenAI's GPT-4o has consistently been the best quality AI, and it is 1/4 the price of the next best, Anthropic's Claude 3 Opus.

The downsides are, the small model is one of the more expensive small models. Still, 14 hours for $1 is a good deal.

Local LLMs

This screenshot shows only 2 of the 4 built-in models, and doesn't show the Custom option.

Local LLMs run on your device. They push the device to its limits, and generally, are slower and lower quality. The smaller the model, the less memory and CPU it requires, but the less powerful it is.

Their advantage is they are free, and because they run on your device, they are completely private, none of your data crosses through the internet.

1. Download

To use a Local LLM, go to Settings > AIs, tap LLMs, choose on of the models under "On Your Device", and press the Get button.

The model will download and show its progress. When it is done, your LLM will be set to that model, and a green checkmark will appear next to it.

2. Test

When the download is done, press Test below the models to make sure it works on your device. You'll see a "Testing..." banner at the top of the screen, followed by "Test Succeded" with a speed measurement.

Test Stuck? Failed?

If the takes longer than 1 minute, it is likely your device is too slow for that model size, and it couldn't load.

If you are on a reasonably powerful device with the smallest model, and don't see a "Test Succeeded" or "Test Failed" message, please use Settings > Contact to let us know.

We should display an error message in that case, but, the AI challenges devices to their limits, sometimes the device can't even display an error message. Having the Telosnex debug logs helps us understand how this occurs on different devices and systems, and how we can help.

Local Models

There are innumerable local LLMs. Telosnex focuses on a curated & tested set of 4 models, and offers a Custom option that makes it easy to use any other model, including multimodal models.

• Phi-3 Mini: Recommended on phones and web. 2 GB in size. Unlike any other model of its size, it can handle search well. Note: web browsers can run this model, but slow: 2 words per second. Android is ~4 words per second. Up to 128,000 tokens, 384 pages.

• Llama 3 1M 8B: Recommended for users on laptops and desktops. 4.9 GB in size. Up to 1 million tokens, 3000 pages.

• Mistral x LLaVa 1.6 7B: Recommended for users on laptops and desktops. This is a multimodal model: it can also process images. 5 GB in size.

• Mixtral 8x7B: Recommended for users on premium laptops and desktops. This model is extremely high quality, it matches ChatGPT in benchmarks. 22 GB in size.

• Custom: Telosnex works with any llama.cpp compatible model, i.e. .gguf files. In practice, this covers every local LLM: all local LLM tools, from Ollama to LM Studio, use llama.cpp.
  • Open .mmproj: Some models come with mmproj files. That means the model is multimodal, and providing the mmproj file allows it to process images.
  • Hugging Face: Opens a search for gguf models on Hugging Face, the largest repository of LLM models. You can browse and download models from there.

Web Search

Search engines take a search query and output web pages. Combining search engines significantly enhances AI answer quality. Telosnex runs an on-device AI to privately identify the best matches, while embedding citations in the answer. This lets Telosnex show you the best sites and skip the spam.

Telosnex offers 2 search engines:

$1 = 1000 searches
Serper describes itself as the world's fastest and cheapest Google search API. It costs 1/10th of a cent per search.

$1 = 111 searches
Brave is a privacy-focused search engine. It costs 9/10th of a cent per search.

Text to Speech

Read text aloud. Telosnex can use text to speech built-in to your device for free, or use cloud-based text to speech for a cost. OpenAI's text to speech offers high quality at a lower cost, while Eleven Labs offers state-of-the-art quality, and many voices, at a higher cost.

Telosnex offers 3 text to speech providers:

Free.
Your device's built-in text to speech. Quality varies by device. You can usually download additional voices from your device's Settings.

$1 = 1 hour
OpenAI's text to speech is high quality, relatively low cost.

$1 = 10 minutes
Eleven Labs offers state-of-the-art quality, and many voices, at a higher cost.

Speech to Text

Free: runs on your device
OpenAI created Whisper, a state-of-the-art voice recognition AI. Telosnex wrote code run the AI on every platform Telosnex runs on.

It is free, and does not require an internet connection.

Tiny (18% error) is recommended for web browsers and phones from before 2021.
Base (15% error) is recommended for most users.
Small (12% error) is best for users on powerful laptops or desktops.

tip

If Whisper isn't set up, when you tap the mic button, the Telosnex Web App will offer to download Tiny. Other Telosnex apps will offer to download Base.

Art

$1 = 25 images
DALL·E 3 is a high-quality art AI built by OpenAI. It is the image generator used by ChatGPT. Telosnex offers direct control & options unavailable in ChatGPT: image size, HD, and Vivid.

$1 = 333 images
Flux Schnell is a high-volume art AI built by Black Forest Labs. It is of excellent quality, roughly the same as DALL·E 3, but it is much faster and much cheaper.

$1 = 20 images
Flux Pro is a high-quality art AI built by Black Forest Labs. It is the highest quality art AI available today, beating even Midjourney in benchmarks. It is also the most expensive.

$1 = 28 images
Stable Diffusion 3 is the latest from pioneers of AI art, Stability.ai. It offers few options, and is the most expensive art AI. However Stable Diffusion 3 has a powerful advantage over DALL·E 3: DALL·E 3 rewrites prompts, often editing out important references, or editing in ahistorical references much like Gemini did, to much controversy. Stable Diffusion 3 does not.

This also lowers quality, relatively, but that is easily addressed with a Script. See Cookbooks > Script Basics & Art Prompter.

API Keys

AI developers often have API keys for providers. Telosnex can use those keys. You don't need to buy credits from Telosnex, instead, the provider will handle billing.

API keys are stored on your device, and your device will use them to access the AI provider's servers directly.

To setup API keys, go to the Settings > AIs, and open API Keys at the bottom.

info

Anthropic does not support calling the API directly from a web app. GitHub issue

If you use the Telosnex web app, Telosnex's servers relay your request to Anthropic's servers.