Skip to content

CURRENT_MODEL_RECOMMENDATIONS.md

Latest commit

 

History

History
106 lines (82 loc) · 5.5 KB

CURRENT_MODEL_RECOMMENDATIONS.md

File metadata and controls

106 lines (82 loc) · 5.5 KB
Raw
title Best Local AI Models July 2026: Current Recommendations
description Up-to-date local AI model recommendations including Qwen 3.5, Gemma 4, gpt-oss, Granite 4.1, and Qwen3-Coder. Refreshed as the Ollama library changes.
keywords best AI models 2026, Qwen 3.5, Gemma 4, gpt-oss, Granite 4.1, Qwen3-Coder, latest local AI models

Best Local AI Models July 2026: Current Recommendations

Last updated after checking the latest Ollama library offerings

Discover the best AI models to install locally right now. Current recommendations across general chat, coding, reasoning, and vision - with the hardware you actually need for each.

My current go-to models

If you're just getting started

  • qwen3.5:4b (~3.4GB) - My #1 recommendation for beginners. Handles text and images, 256K context
  • gemma3:4b (~3.3GB) - Google's small workhorse, very reliable
  • granite4.1:3b - IBM's small model, strong at instruction following and tool calling
  • qwen3.5:2b (~2.7GB) - When memory is really tight

For coding

  • qwen3-coder:30b - A mixture-of-experts model that punches way above the speed you'd expect from "30B". This is my daily driver if the hardware allows
  • qwen3.5:9b (~6.6GB) - The best all-rounder that still fits comfortably in 16GB
  • devstral-small-2:24b - Built specifically for agentic coding across multiple files
  • qwen2.5-coder:7b - Older, but still a solid pick on modest hardware

If you've got a powerful machine

  • gpt-oss:20b - OpenAI's open-weight model. Great reasoning for its size
  • gemma4:31b (~20GB) - Google's current flagship for a single GPU. Vision, audio, tool calling
  • qwen3.5:35b (~24GB) - MoE with ~3B active parameters, so it's faster than the size suggests
  • qwen3.6:27b - Newest Qwen release, geared toward agentic coding

Vision and specialist models

  • qwen3-vl:8b - Solid image understanding without a huge download
  • minicpm-v4.5:8b - Strong at images and video frames
  • embeddinggemma:300m / nomic-embed-text - Embeddings for search and RAG

What's still worth using vs what's getting old

Models that are current and good

  • qwen3.5 series (0.8b, 2b, 4b, 9b, 27b, 35b - multimodal, 256K context)
  • qwen3.6 (27b, 35b - newest Qwen, agentic coding focus)
  • gemma4 (e2b, e4b, 12b, 26b, 31b - vision, audio, tools)
  • gemma3 (1b, 4b, 12b, 27b - still a great default)
  • gpt-oss (20b, 120b - OpenAI's open-weight reasoning models)
  • granite4.1 (3b, 8b, 30b - Apache 2.0, enterprise friendly)
  • qwen3-coder (30b, 480b) and qwen3-coder-next
  • devstral-small-2:24b (agentic coding)
  • ministral-3 (3b, 8b, 14b - built for edge devices)
  • lfm2.5:8b (fast tool calling on consumer hardware)
  • nemotron-3-nano (4b, 30b)
  • olmo-3.1:32b (fully open training data, great for learning how this works)

Models you should probably replace

  • llama3.2:3b → qwen3.5:4b or gemma3:4b (Llama 3.2 is nearly two years old now)
  • llama3.1:8b / llama3.3:70b → qwen3.5:9b or gpt-oss:20b
  • qwen2.5:7b → qwen3.5:9b
  • phi3.5:3.8b → phi4-mini:3.8b, or just move to qwen3.5:4b
  • granite-code / granite3-dense → granite4.1
  • codellama:7b → qwen3-coder:30b or qwen2.5-coder:7b
  • gemma2:9b → gemma3:12b or gemma4:12b
  • deepseek-r1:7b → still fine, but gpt-oss:20b or qwen3.5 with thinking enabled is better if it fits

Things worth knowing about (2026)

  • Mixture-of-experts (MoE) models are everywhere now. Tags like 35b-a3b mean 35B total parameters but only ~3B active per token. You still need memory for the full 35B, but the speed feels like a much smaller model
  • Thinking/reasoning modes are built in. Many models now toggle a reasoning mode instead of shipping a separate "reasoning model"
  • Multimodal is the default. Qwen 3.5 and Gemma 4 handle images out of the box; Gemma 4 also does audio
  • New quantization formats - you'll see mxfp4, nvfp4, int4, and QAT tags alongside the familiar q4_K_M. QAT versions are quantization-aware trained, so they hold up better at small sizes
  • Cloud-tagged models run on Ollama's servers, not yours. Convenient, but not private and not offline

Just tell me what to download

Want to try AI for the first time? → ollama run qwen3.5:4b

Need help with programming? → ollama run qwen3-coder:30b (or qwen2.5-coder:7b on lighter hardware)

Work in multiple languages? → ollama run qwen3.5:9b

Got a beast machine and want the best? → ollama run gpt-oss:120b or ollama run qwen3.5:122b

Want something balanced and reliable? → ollama run gemma4:12b

Need lightweight but capable? → ollama run qwen3.5:2b or ollama run granite4.1:3b

Want the latest reasoning capabilities? → ollama run gpt-oss:20b

Need it to look at images? → ollama run qwen3-vl:8b

Some practical advice

  1. Don't start with the biggest model - Begin with 3B-9B, upgrade later if needed
  2. Check the download size, not the parameter count - An MoE model tagged 35b-a3b is fast, but it still needs ~24GB of memory
  3. Watch your resources - ollama ps tells you whether a model landed on the GPU or fell back to CPU
  4. Use specialized models - Coding models really are better at coding, and reasoning models are worth the extra wait on hard problems
  5. Check back regularly - New models drop constantly, and some are genuinely better

I try to keep this updated as I test new models and see what's actually working well in practice. Last checked: July 2026