You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
Copy file name to clipboardExpand all lines: docs/ai-chat.md
+69-10Lines changed: 69 additions & 10 deletions
Display the source diff
Display the rich diff
Original file line number
Diff line number
Diff line change
@@ -25,9 +25,9 @@ Alternatively, you can run AI models locally so that your data never leaves your
25
25
26
26
### Hardware for Local AI Models
27
27
28
-
Local models are also fairly accessible. It's possible to run smaller models at lower speeds on as little as 8 GB of RAM. Using more powerful hardware such as a dedicated GPU with sufficient VRAM or a modern system with fast LPDDR5X memory offers the best experience.
28
+
Local models are also fairly accessible. They can run on most PCs and some high-end smartphones. It's possible to run smaller models at lower speeds on as little as 8 GB of RAM. Using more powerful hardware such as a dedicated GPU with sufficient VRAM or a modern system with fast LPDDR5X memory offers the best experience.
29
29
30
-
LLMs can usually be differentiated by the number of parameters, which can vary between 1.3B to 405B for open-source models available for end users. For example, models below 6.7B parameters are only good for basic tasks like text summaries, while models between 7B and 13B are a great compromise between quality and speed. Models with advanced reasoning capabilities are generally around 70B.
30
+
LLMs can usually be differentiated by the number of parameters, which can vary between 1.3B to 405B for open-source models available for end users. For example, models below 3B parameters are only reliably good for simple lingustic tasks such as summarization, models above 4B start understanding context and have a good knowledge of the world. From 8B, they can exercise basic reasoning. Models with advanced reasoning capabilities start at around 30B parameters.
31
31
32
32
For consumer-grade hardware, it is generally recommended to use [quantized models](https://huggingface.co/docs/optimum/en/concept_guides/quantization) for the best balance between model quality and performance. Check out the table below for more precise information about the typical requirements for different sizes of quantized models.
33
33
@@ -45,16 +45,75 @@ There are many permissively licensed models available to download. [Hugging Face
45
45
46
46
To help you choose a model that fits your needs, you can look at leaderboards and benchmarks. The most widely-used leaderboard is the community-driven [LM Arena](https://lmarena.ai). Additionally, the [OpenLLM Leaderboard](https://huggingface.co/spaces/open-llm-leaderboard/open_llm_leaderboard) focuses on the performance of open-weights models on common benchmarks like [MMLU-Pro](https://arxiv.org/abs/2406.01574). There are also specialized benchmarks which measure factors like [emotional intelligence](https://eqbench.com), ["uncensored general intelligence"](https://huggingface.co/spaces/DontPlanToEnd/UGI-Leaderboard), and [many others](https://nebuly.com/blog/llm-leaderboards).
47
47
48
+
#### Recommended models
49
+
50
+
Below is a table of good models to start with.
51
+
52
+
|Developer|Model name| Available sizes| Strengths | Weaknesses | Censorship |
**Edge Gallery** is an AI client developped by Google that features Google's latest open-weights models. It runs fully locally, but downloading the preloaded models require a Hugging Face account as you will need to [accept Google's Acceptable Use policy](https://github.com/google-ai-edge/gallery/wiki/5.-Model-Management#hugging-face-login--license-acknowledgement) to access the model files. You can import models you downloaded, but the app only accepts the [LiteRT](https://github.com/google-ai-edge/litert) format, which isn't widely used.
77
+
78
+
The preloaded models are performant and can analyse an image. However, you might not be able to have long conversation or analyse long documents, as the app restricts the maximum context length to an ammount [multiple times smaller](https://github.com/google-ai-edge/gallery/issues/183) than what the AI model can normally support.
**PocketPal** is an AI client available on both iOS and Android. It focus on customisability, ease-of use, and above all on small language models (SLMs), allowing you to run AI models on basically any Android phone. Keep in mind however that SLMs aren't useful in many domains.
100
+
101
+
Other models can easily be imported, either from a Hugging Face repo link, of from a gguf file directly. Models can also be customised, and the client also offers "Pals", which are describe as "personalized AI assistants with different personalities: but are essentially presets of a model and a system prompt. You can chose between an assitant pal and a roleplay Pal.
0 commit comments