ZwZ2-9B: novidades, hardware necessário, como rodar grátis

ZwZ2-9B lifts overall accuracy from 51.63 % to 56.59 % across 25 benchmarks. It does this without any extra supervised fine-tuning.

ZwZ2-9B is a 9.4 billion-parameter model built on the Qwen3.5-9B backbone. The model is trained on whole images, not cropped patches, and can handle a context window of 256 K tokens. It is released by InclusionAI and available for free on Hugging Face. The model card lists the size, the base model, and the training approach, giving hobbyists a clear picture of what they are downloading.

The core novelty is the EvidenceConnect pipeline. It builds verified questions that span three to seven related image regions, using privileged zoomed views just for data construction. Answers are checked through cross-model agreement and drop-one-region tests, then turned into four-option tasks. Training uses the GRPO method and skips any additional supervised fine-tuning stage, keeping the process simple and efficient.

According to the model card, ZwZ2-9B was evaluated on twenty-five benchmarks covering fine-grained perception, spatial reasoning, long-context understanding, multimodal agents, and general visual reasoning. Across all of them the model improves over its Qwen3.5-9B base, raising the average accuracy from 51.63 % to 56.59 %, a gain of 4.96 points. The improvement is consistent on every benchmark, showing the method’s broad impact.

The model’s strengths lie in its enhanced ability to connect visual evidence across regions, which translates into better fine-grained perception and stronger spatial and long-context reasoning. However, the model card warns that it can still hallucinate details, misread text, and make counting or spatial errors. Performance also depends on image resolution, context length, and prompt format, so careful prompt engineering helps mitigate these limits.

ZwZ2-9B is a good fit for hobbyists who run AI models at home and need visual reasoning capabilities without paying for an API. The model can run on a modest RTX 4060, making it accessible to many desktop setups. If you want a free, open-weight model that improves visual understanding over a strong base, ZwZ2-9B is worth a try.

O que você precisa para rodá-lo em casa? Com 4 bits, ele ocupa cerca de 8 gigabytes de memória. A menor máquina doméstica que o comporta por inteiro é RTX 4060. Em RTX 3060 12 GB com 32 gigabytes de RAM, ele escreve cerca de 45 tokens por segundo. Em RTX 4090 com 64 gigabytes de RAM, ele escreve cerca de 116 tokens por segundo.

There is no free API for it yet. The weights are open: you can download them from Hugging Face, under the license written in its model card. Arena has not ranked it yet: we add the score as soon as the votes arrive.

Compare ZwZ2 9B with every free model, and check what your own GPU can run, on FreeLLMArena.