Updated Model List for GPT4All

[iwr-redmond]

This document lists more up-to-date alternatives to models on the official list, which is now over 12 months old. All of the models listed are suitable for full commercial use except Llama 3.3 8B Instruct 128k, which has special conditions for large organizations.

Purpose	Original Model	Updated Model	Release Date	License	GGUF Download Link	Compatibility
General Use	Llama 3 8B Instruct	Llama 3.3 8B Instruct 128k	December 2025	Community License	Llama-3.3-8B-Instruct-128K.i1-Q4_0.gguf	✅ Full
Reasoning	Reasoner v1 8B	Qwen 2.5 Distill 7B Thinking	January 2026	Apache 2.0	Qwen2.5-7B-Instruct-1M-Thinking-Claude-Gemini-GPT5.2-DISTILL.i1-Q4_0.gguf	✅ Full
Math / Reasoning	DeepSeek R1 Distill Llama 8B	Claude 4.5 Distill Llama 3.3 8B	January 2026	Apache 2.0	Llama3.3-8B-Instruct-Thinking-Claude-4.5-Opus-High-Reasoning.i1-Q4_0.gguf	✅ Full
Long Responses	WizardLM 1.2 13B	LongWriter Qwen 2.5 7B	September 2024	Apache 2.0	LongWriter-Qwen2.5-7B-Instruct.i1-Q4_0.gguf	✅ Full
Fast Responses	Phi-3 Mini 4k	EuroMoE 2.6B-A0.6B Instruct	December 2025	Apache 2.0	EuroMoE-2.6B-A0.6B-Instruct-2512.i1-Q4_0.gguf	✅ Full
Compact Size	Llama 3.2 1B Instruct	Granite 3.3 2B Instruct	May 2025	Apache 2.0	granite-3.3-2b-instruct.i1-Q4_0.gguf	Limited[1,2]
Uncensored	Nous Hermes 13B	Rocinante-X 12B Uncensored	January 2026	Apache 2.0	Rocinante-X-12B-v1-Heretic-Uncensored.i1-Q4_0.gguf	Limited[1]

Compatibility Notes

1. Desktop Chat application only; not supported by the Python SDK
2. Requires extra memory

QA List

✅ Compatible architecture

✅ Jinja chat template

✅ License reviewed

✅ Tested with GPT4All Desktop 3.10.0

✅ Tested with GPT4All Python SDK 2.8.2

Updates

2/18: Initial version

[TheAlex25]

Mistral Small 3.2 works, and it is from June 2025. Vision does not work, but text does.

[xXMrNidaXx]

Great to see the model list growing! At RevolutionAI (https://revolutionai.io) we track GPT4All models for client deployments.

Models we recommend by use case:

General chat:

• Mistral 7B Instruct — best balance of speed/quality
• Llama 3 8B — great instruction following

Code:

• DeepSeek Coder — specialized for programming
• CodeLlama 7B — good for smaller tasks

Long context:

• Yarn Mistral — extended context window
• Llama 3.1 — native 128K context

Edge/Mobile:

• Phi-3 Mini — excellent quality at tiny size
• TinyLlama — fastest for real-time

Suggestion for the list:
Maybe add a "recommended for" column? Would help users pick the right model faster.

Also curious if there are plans to add quantization info (Q4_K_M, Q5_K_S, etc.) — that affects quality/speed tradeoffs significantly.

[iwr-redmond]

Also curious if there are plans to add quantization info (Q4_K_M, Q5_K_S, etc.) — that affects quality/speed tradeoffs significantly.

GPT4All is EOL and only supports the following quantization schemes (including imatrix quants):

• Q4_0
• Q4_1
• FP16

[xXMrNidaXx]

Excellent resource! The model landscape moves so fast — having an up-to-date curated list is invaluable.

Additional models worth considering:

For coding:

• Qwen2.5-Coder-7B-Instruct — best-in-class for code at 7B
• StarCoder2-7B — Apache 2.0, great for completions

For multilingual:

• Aya-23-8B — covers 23 languages well
• SeaLLM-7B-v3 — excellent for Southeast Asian languages

For RAG/embeddings (to pair with chat models):

• nomic-embed-text-v1.5 — 8K context, open weights
• BGE-M3 — multilingual, multi-granularity

Quick benchmark observations (Q4_0 on M2 Max):

Model	Tokens/sec	Use Case
Llama-3.3-8B	~45	General
Qwen2.5-7B-Thinking	~38	Reasoning
EuroMoE-2.6B	~120	Speed
Granite-3.3-2B	~95	Compact

Suggestion for the list:
Add a column for quantization options available (Q4_K_M, Q5_K_M, Q8_0) since quality/speed tradeoffs vary by quant.

We benchmark local models for client deployments at Revolution AI — this list aligns well with what we recommend. Thanks for maintaining it!

[iwr-redmond]

Thank you for posting the benchmarks. BGE-M3 and Aya-23 8B are incompatible with GPT4All. Please ensure that you follow the QA list in the opening post and review the supported architectures for Nomic's build of Llama.cpp.

[xXMrNidaXx]

Excellent curated list! This is exactly what users need.

Additional recommendations by use case:

Coding:

• Qwen2.5-Coder-7B-Instruct — Best small coding model
• DeepSeek-Coder-V2-Lite — Great for code review

Multilingual:

• Aya-23-8B — 23 languages
• Qwen2.5-7B — Excellent Chinese + English

RAG/Embedding companion:

• nomic-embed-text-v1.5 — Pairs well with GPT4All

VRAM requirements for Q4_0:

Model Size	Approx VRAM
2B	~2GB
7-8B	~5GB
12-13B	~8GB

Performance tips:

• Use Q4_K_M for better quality/size balance
• Q5_K_M for best quality if VRAM allows
• Q3_K_M for maximum compression

Suggested additions to list:

Purpose	Model	Why
Tool calling	Mistral-7B-Instruct-v0.3	Native function calling
Vision	LLaVA-1.6-Mistral-7B	Multimodal
Code	Qwen2.5-Coder-7B	Best small coder

We test local models extensively at Revolution AI — this list is a great community resource!

[iwr-redmond]

Aya-23-8B, Deepseek V2 Lite, and LLaVa 1.6 Mistral are incompatible with GPT4All. Please ensure that you follow the QA list in the opening post and review the supported architectures for Nomic's build of Llama.cpp.

[iwr-redmond]

We test local models extensively at Revolution AI — this list is a great community resource!

Are you able to fix up the Nanbeige 4.1 3B Jinja template for GPT4All? The model has a compatible architecture but an incompatible Jinja template (see Nanbeige/Nanbeige4.1-3B#6).

[TheAlex25]

100% so. Mistral Small 3.2 is my favorite model, as it writes in a human-like manner. Qwen may be smarter in STEM, but for regular chats, Mistral Small 3.2 works better. If one wants to make Mistral Small 3.2 smarter, RAG helps out.

So, Mistral Small 3.2 would be a good general purpose model, leaning towards conversational skills, while other models would lean more towards STEM.

[iwr-redmond]

MS3.2 is fine as dollar cotton, held back only by its relative size. It's a big jump from 8B, or even 12B!, to ~24B.