Table of Content
ollama v0.6.6 is released! The reasoning ability is doubled, and the download speed is increased by 50%. Which one is stronger than vLLM/LMDeploy?
Updated on:June-30th-2025
Recommendation
Ollama v0.6.6 is coming, the reasoning capability is doubled, the download speed is increased by 50%, a new choice for AI developers!
Core content:
1. Introducing two new models, Granite 3.3 and DeepCoder, to enhance reasoning and code generation capabilities
2. Download speed is significantly improved, memory leaks are fixed, and the operation is more stable
3. API and compatibility improvements, ease of use, reasoning speed and memory optimization are comprehensively improved
Yang Fangxian
Founder of 53AI/Most Valuable Expert of Tencent Cloud (TVP)
Ollama v0.6.6 major update: stronger reasoning, faster download, more stable memory
Attention AI developers! Ollama v0.6.6 is officially released, bringing many major optimizations, including new model support, faster download speed, memory leak fixes, etc., making local large model inference more efficient and stable!
? Core update highlights
1. Two new models are launched
• Granite 3.3 (2B & 8B): 128K ultra-long context , optimized instruction following and logical reasoning capabilities, suitable for complex task processing. • DeepCoder (14B & 1.5B): A completely open source code model with performance comparable to O3-mini. Developers can deploy high-quality code generation AI at low cost!
2. Download speed is greatly improved
• Experimental new downloader : via OLLAMA_EXPERIMENT=client2 ollama serveEnable, download faster and more stable!• Safetensors import optimization : ollama createSignificant performance improvements when importing models.
3. Critical BUG fixes
• Gemma 3 / Mistral Small 3.1 memory leak issue fixed , running more stable. • OOM (out of memory) issue optimization , reserve more memory at startup to avoid crashes. • Safetensors import data corruption issue fixed to ensure model integrity.
4. API and compatibility improvements
• Supports tool function parameter type arrays (such as string | number[]), the API is more flexible.• OpenAI-Beta CORS header support for easy front-end integration.
? Ollama vs. vLLM vs. LMDeploy: Who is the king of local deployment?
| Comparison Dimensions | Ollama v0.6.6 | vLLM | LMDeploy |
| Ease of use | |||
| Inference speed | |||
| Memory optimization | |||
| Model support | |||
| Applicable scenarios | Personal development/lightweight application | Highly concurrent production environment | Enterprise-level real-time conversation/edge computing |
Related PostsView More
Enhanced fine-tuning is coming! How to make AI truly "understand" human needs
In-depth exploration of the secrets of fine-tuning large models! At a time when AI is developing rapidly, how to make models better understand human needs is the key. The emergence of reinforcement fine-tuning brings new hope. It combines human feedback with reinforcement learning to optimize model output. The article introduces model fine-tuning methods in detail, such as the principles of...
Treasure Discovery: Sim Studio, an open source tool that makes AI workflow construction simple
Discover a treasure! Sim Studio is an open source tool that makes it easy to build workflows for large language models. It is a powerful and user-friendly open source platform that provides a visual interface and lowers the threshold for use. Whether verifying AI application ideas or building complex workflows, it can be done efficiently. It supports multiple deployment methods, including...
What? Just by speaking, you can let a large model help you organize your computer desktop?!
Want to let the big language model help you organize your computer desktop easily? This article explains it in detail! Based on the MCP protocol, use Golang to build an intelligent cleaning system, and combine the big language model to achieve automated operations. It not only introduces the MCP architecture design and communication process, but also the details of Server implementation. New...
