According to CoinMeta, NVIDIA released Nemotron 3.5, claiming it to be the best open-source model in the United States, despite scoring nearly 20 points behind Xiaomi. This model has a total of 120 billion parameters, with only 12 billion activated at a time, and supports up to 1 million token contexts. It combines text reasoning, visual understanding, and tool invocation. The new model uses a hybrid architecture of mamba and transformer, and improves inference efficiency through multi-token predictions and four-bit quantization. With 100,000 token inputs, the output speed is about 630 token per second, which is approximately 5 times that of Inkling Small models with similar capabilities. NVIDIA also stated that using 4 B200 GPU to process 10,000 programming tasks in batches results in a speed that is up to 4 times faster than some other open-source models in the same category. In the swe-bench pro tests, this model scored 26.5, ranking first among American open-source models, but it still lags behind Xiaomi's Mimo-v2.6-pro (which scored 46 points).