Skip to content

CURRENT_MODEL_RECOMMENDATIONS.md

Latest commit

 

History

History
138 lines (105 loc) · 7.04 KB

CURRENT_MODEL_RECOMMENDATIONS.md

File metadata and controls

138 lines (105 loc) · 7.04 KB
title Best AI Models August 2025: Latest Local AI Recommendations
description Get the latest AI model recommendations for 2025 including Llama 3.3, DeepSeek R1, SmolLM2, and Granite3. Updated monthly with current best practices.
keywords best AI models August 2025, Llama 3.3, DeepSeek R1, SmolLM2, latest AI models, AI model recommendations 2025

Best AI Models August 2025: Current Recommendations for Local Installation

Last updated after checking the latest Ollama library offerings

Discover the best AI models to install locally in 2025. Get current recommendations for Llama 3.3, DeepSeek R1, Qwen 2.5, and other top-performing models for different use cases.

🚀 My current go-to models

If you're just getting started

  • llama3.2:3b - Still my #1 recommendation for beginners
  • gpt-oss - OpenAI's new open models - incredibly capable with reasoning and function calling
  • smollm2:1.7b - Microsoft's lightweight model that punches above its weight
  • qwen3:1.7b - New entry from Alibaba, very capable for its size
  • phi3.5:3.8b - Microsoft's solid alternative, reliable choice

For coding stuff

  • deepcoder:14b - New open-source coding champion, performs at o3-mini level
  • opencoder:8b - Bilingual coding model supporting English and Chinese
  • starcoder2:7b - Transparently trained open code model
  • codestral - Mistral's first dedicated code generation model
  • codegemma - Google's lightweight coding specialist
  • stable-code:3b - Efficient specialist that rivals larger models
  • qwen3-coder:7b - Latest coding model from the Qwen 3 family
  • deepseek-coder-v2:16b - Amazing if you've got the hardware to run it

For advanced reasoning and complex tasks

  • gpt-oss - OpenAI's open models with configurable reasoning effort
  • phi-4:14b - Microsoft's new reasoning powerhouse rivaling much larger models
  • qwq:32b - Reasoning-focused model, excellent for complex problems
  • tulu3 - Allen Institute's leading instruction-following model
  • deepseek-v3 - Massive 671B parameter model (37B active) - cutting edge
  • olmo2:13b - Competitive with Llama 3.1, great performance

If you've got a powerful machine

  • athene-v2:72b - Excellent for mathematics and technical tasks
  • qwen3:32b - Really impressive performance across the board
  • llama3.1:70b - Still a powerhouse for general tasks
  • llama3.3:70b - Meta's latest large model
  • granite3.2 - IBM's updated models with 128K context
  • mistral-small-3.1 - Great balance with vision capabilities

For vision and multimodal tasks

  • mistral-small-3.1 - Latest with vision understanding and 128K context
  • llama3.2-vision:11b - Text + image understanding in one model
  • llava:13b - Excellent for visual Q&A and image description

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

✅ Models that are still current and good

  • llama3.2:3b (my daily driver)
  • llama3.3:70b (Meta's newest, if you can run it)
  • phi3.5:3.8b (reliable and efficient)
  • qwen2.5 series (7b, 14b, 32b - all solid)
  • gemma2:9b, gemma2:27b (Google's offerings, very capable)
  • qwen2.5-coder:7b (still the best general coding model)
  • deepseek-coder-v2:16b (if you can run it)
  • deepseek-r1:7b (new reasoning model, really impressive)
  • mistral-nemo:12b (excellent quality)
  • granite3-dense:8b (IBM's improved version)
  • smollm2:1.7b (surprisingly good for its size)

🔄 Models you should probably replace

  • granite-code:8b → Switch to granite3-dense:8b (newer version)
  • codellama:7b → Switch to qwen2.5-coder:7b or starcoder2:7b
  • llama3.1:8b → Consider llama3.2:8b or llama3.3:70b for improvements

🆕 What's new in August 2025?

The AI world moves fast! Here's what just dropped:

  • GPT-OSS - OpenAI released their first open-weight models with incredible reasoning capabilities, function calling, and configurable thinking effort. Already 84.1K pulls in just 5 hours!
  • Phi-4 - Microsoft's new 14B reasoning model that rivals much larger models in complex reasoning tasks
  • DeepCoder - A fully open-source 14B coding model performing at the level of o3-mini
  • Mistral Small 3.1 - Adds state-of-the-art vision understanding to Mistral's capabilities
  • Tülu 3 - Allen Institute's leading instruction-following model family
  • Qwen 3 family - Alibaba's massive update with models from 0.6B to 235B parameters, including MoE variants
  • Athene-V2 - 72B parameter model excelling in mathematics and technical tasks
  • Granite 3.2 - IBM's updated models with 128K context and improved reasoning
  • OpenCoder - Bilingual coding model supporting both English and Chinese
  • Command R7B - Cohere's model with advanced Arabic language capabilities
  • Dolphin - Uncensored instruct-tuned models for flexible applications
  • StarCoder2 - Updated transparently trained open code models
  • CodeGemma - Google's lightweight coding specialist
  • Bespoke-Minicheck - Factuality checking model to detect hallucinations

💡 My honest take on what to actually use

I've been testing these models for weeks. Here's what I'm actually running:

For most people: Still start with llama3.2:3b for reliability, but if you want to try the cutting edge, gpt-oss is genuinely impressive for an open model.

For coding: deepcoder:14b has blown me away - it's performing at o3-mini level for coding tasks. If that's too big, stable-code:3b remains excellent.

For reasoning: phi-4:14b and gpt-oss are both genuinely good at thinking through complex problems. The configurable reasoning effort in GPT-OSS is particularly cool.

For vision tasks: mistral-small-3.1 now has vision capabilities alongside its text skills, making it a great all-in-one option.

For power users: athene-v2:72b is incredible for mathematical and technical work if you have the hardware to run it.

🎯 Just tell me what to download

🆕 Want to try AI for the first time? → ollama run llama3.2:3b

💻 Need help with programming? → ollama run qwen2.5-coder:7b

🌍 Work in multiple languages? → ollama run qwen2.5:7b

🔥 Got a beast machine and want the best? → ollama run llama3.3:70b or ollama run deepseek-r1:32b

⚖️ Want something balanced and reliable? → ollama run mistral-nemo:12b

🪶 Need lightweight but capable? → ollama run smollm2:1.7b

🧠 Want the latest reasoning capabilities? → ollama run deepseek-r1:7b

💡 Some practical advice

  1. Don't start with the biggest model - Begin with 3B-7B, upgrade later if needed
  2. Try the new reasoning models - deepseek-r1 series is genuinely impressive for complex problems
  3. Watch your resources - Keep an eye on RAM/VRAM usage, especially at first
  4. Use specialized models - Coding models really are better at coding, reasoning models excel at complex problems
  5. Check back regularly - New models drop frequently, 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: August 2025