Open-weight models

An open-weight model publishes its weights. That is separate from being free to call. Qwen3 32B is open-weight and $0.08 per million input tokens on OpenRouter. Claude Sonnet 5 is not in this group.

A practical shortlist

• DeepSeek V4 Pro and V4 Flash, when you want a million-token window and a published MoE card.
• Qwen3 32B and Qwen3 235B A22B, when you want dense or MoE Qwen weights and a 131K window.
• Qwen3.5 9B, when the prompt includes an image and the budget is small.
• GLM 5.2, when the agent run is long and you can pay about $2 per million output tokens.
• Gemma 4 and Nemotron :free slugs, when the call must be $0 and quality can vary.

OpenRouter’s June 2026 article, “the open-weight models that matter,” is their own selection for that month. Models have shipped since. Treat the article as a dated essay and the models API as the price list.

If you need to fine-tune or run offline, the OpenRouter slug does not hand you the files. Follow the model card’s weights link. The API price then becomes irrelevant and your GPU cost replaces it.