Omnivision-968M: Vision Language Model with 9x Tokens Reduction for Edge Devices