Hey HN! We’re Kasey and Andy from Osmosis (https://osmosis.ai/). We’ve been playing with MCP recently and wanted to share a lightweight open source model that can connect any MCP client* to any MCP server! Check it out here: https://huggingface.co/osmosis-ai/osmosis-mcp-4b
Right now, the only models that consistently work well for MCP are large and closed-source (e.g. 3.7 Sonnet, Gemini 2.5). Other models struggle with tool-calling consistency, and in particular, there’s a lack of options that can run locally.
We used Dr. GRPO to train Qwen3-4B (w/ VeRL + SGLang for multi-turn tool-calling training) for this purpose, and we were able to get performance parity with Gemini 2.5 Pro on relevant benchmarks like GSM8K (https://i.imgur.com/4RXq2Pm.png).
And since the model is open source and lightweight, that means you can also do further fine-tuning / training that’s fully local and customized to your specific needs.
Let us know what you think, or if there’s anything we can answer!
* Any client that supports Qwen3 models (i.e. it works with OpenRouter, local deployments with Ollama, etc.)
Right now, the only models that consistently work well for MCP are large and closed-source (e.g. 3.7 Sonnet, Gemini 2.5). Other models struggle with tool-calling consistency, and in particular, there’s a lack of options that can run locally.
We used Dr. GRPO to train Qwen3-4B (w/ VeRL + SGLang for multi-turn tool-calling training) for this purpose, and we were able to get performance parity with Gemini 2.5 Pro on relevant benchmarks like GSM8K (https://i.imgur.com/4RXq2Pm.png).
And since the model is open source and lightweight, that means you can also do further fine-tuning / training that’s fully local and customized to your specific needs.
Let us know what you think, or if there’s anything we can answer!
* Any client that supports Qwen3 models (i.e. it works with OpenRouter, local deployments with Ollama, etc.)