Best Open-Source LLMs for Phones in 2026

For most phones, the best open-source LLM in 2026 is Qwen 3.5 4B: about 2.7 GB, runs on phones with 3 GB of RAM or more, understands images, and has a thinking mode for harder questions. If your phone has 6 GB or more, Qwen 3.5 9B gives noticeably better answers. For older or low-memory phones, Gemma 4 E2B is the most capable small option. For documents and screenshots on a high-end phone, GLM 4.6V Flash stands out.

All of these are open-weight models: their makers publish the files, so you can download them once and run them offline on your own phone.

Quick comparison #

Sizes and RAM figures below are for the compressed (quantized) phone builds in the Personal LLM catalog.

Every model in this list can take an image as input when paired with its vision add-on file (roughly 195 MB to 1 GB extra).

Qwen 3.5 (0.8B, 4B, 9B): best overall #

Qwen 3.5 arrived in February 2026 from Alibaba’s Qwen team under the Apache 2.0 license. According to its model card, it covers 201 languages and dialects, handles images natively, and is designed to reason step by step before answering.

• Qwen 3.5 4B is the sweet spot. It’s quick on everyday phones and good at writing, explaining, summarizing and short code. With thinking mode on, it works through math and logic step by step.
• Qwen 3.5 9B is the strongest general model that fits on a phone today. It needs 6 GB of RAM, and more to run comfortably. It’s slower, but more careful with facts and reasoning.
• Qwen 3.5 0.8B is for phones that can’t run anything else. It’s very fast, but answers are simple and it makes more mistakes.

Wrong for: anyone who needs the fastest possible replies on a low-end phone, where Gemma 4 E2B may feel snappier.

Gemma 4 E2B and E4B: best for small and mid-range phones #

Google DeepMind built the Gemma 4 “E” models for phones. The E stands for effective parameters: E2B has 2.3 billion effective parameters but uses per-layer embeddings that sit outside the main computation, which keeps it light for its quality (model card). E4B has about 4.5 billion effective parameters. The phone builds are quantization-aware trained, meaning the model was trained to hold up well when compressed.

Gemma 4 was trained on more than 140 languages, which makes it a strong choice for non-English chat. See offline AI in other languages.

• E2B is the most capable model for 3 to 4 GB phones.
• E4B gives better answers with hybrid thinking for harder questions, and wants 5 GB of RAM or more.

Wrong for: phones with 8 GB or more, where Qwen 3.5 9B will usually answer better.

Ministral 3 3B: best for long documents #

Mistral AI released Ministral 3 in December 2025 under Apache 2.0. The 3B version pairs a 3.4B language model with a small vision encoder and supports a 256k-token context window (model card). On a phone you won’t use anything close to 256k tokens, since memory runs out first, but the model is built to handle long inputs gracefully, which helps with documents and screenshots.

Wrong for: phones with 6 GB of RAM or more, where Qwen 3.5 4B or 9B are the stronger general-purpose picks.

GLM 4.6V Flash: best vision model for high-end phones #

Zhipu AI’s GLM 4.6V Flash is a 9B vision-language model released under the MIT license. Its model card highlights document understanding (charts, tables and figures) and reading user interfaces from screenshots. Zhipu also notes its limits: pure text question-answering is weaker than its vision skills, and it can overthink or repeat itself.

Wrong for: phones under 8 GB of RAM, and text-only chat, where Qwen 3.5 9B is the better pick.

How do you choose? #

1. Check your RAM. It’s the hard limit. How much RAM you need explains why and how to find yours.
2. Pick the biggest model with headroom. A model that just barely fits will be slow and may get closed when you switch apps. Aim for about 2 GB of spare RAM above the model’s minimum.
3. Try two. Ask each the same three questions and keep the one that reads better.
4. Match the task. Photos and documents: GLM 4.6V Flash or Ministral 3. Languages: Gemma 4 or Qwen 3.5. General chat: Qwen 3.5.

What about models not on this list? #

Plenty of other open models exist, and many have GGUF builds on Hugging Face. If you want one of them, apps that accept custom GGUF files let you add it by URL. Our guide to adding a custom GGUF model to your phone covers picking the right file and quantization.

Running these models on your phone #

Personal LLM includes all seven models above in its catalog, with a “Fits your device” badge that reads your phone’s real RAM before you download. Qwen 3.5 4B is the recommended first download. You can add image support to any of them, switch between them mid-conversation, and compare their speed on your hardware with the built-in benchmark. It runs on iOS 15.1 or later and Android 7 or later, with no account, and it’s free.

Frequently asked questions #

What is the best LLM to run on a phone? #

For most phones, Qwen 3.5 4B offers the best balance of quality, speed and size. On phones with 6 to 8 GB of RAM or more, Qwen 3.5 9B gives better answers, and on low-memory phones Gemma 4 E2B is the strongest small option.

Are these models really free? #

Yes. Qwen 3.5, Gemma 4 and Ministral 3 are released under the Apache 2.0 license, and GLM 4.6V Flash under MIT. You can download and use them at no cost.

Which phone LLM is best for images? #

Every model here supports images with its vision add-on. GLM 4.6V Flash is the strongest at documents, charts and screenshots, but needs 8 GB of RAM. On smaller phones, Qwen 3.5 4B and Ministral 3 3B handle photos well.

Can a phone run Llama or other models? #

If an app accepts custom GGUF files, you can load most open models that have a phone-sized GGUF build, as long as your phone has enough RAM. Performance and quality vary, so start with a well-known publisher’s file.