Gemma 4 MTP and DeepSeek V4 Flash Bring Local LLMs to New Heights in 2026

Google and DeepSeek have released new large language models in late April 2026. Google's Gemma 4 introduces Multi-Token Prediction (MTP) for faster inference. DeepSeek launched V4 Flash and V4 Pro, with V4 Flash optimized for local deployment.

Gemma 4 MTP delivers up to 2x speedup

Gemma 4 with MTP achieves a speedup of 1.24x to 1.74x for single requests and over 2x for four parallel requests on DGX Spark. DeepSeek V4 Flash has 284 billion total parameters with 13 billion active, a 1 million token context, and an MIT license. The model is designed to run on consumer hardware.

Developer antirez, creator of Redis, released DwarfStar 4, a Metal-native inference engine for DeepSeek V4 Flash. It uses 2-bit quantization and disk-based KV cache to run the 86.7 GB model on MacBooks with 128 GB of RAM. Generation speed reaches 31.29 tokens per second with short context and 23.22 tokens per second with long context. DwarfStar 4 supports OpenAI API and Claude Code integration.

DeepSeek V4 Flash and V4 Pro were released globally on April 24, 2026. According to PC Watch, Gemma 4 with MTP can achieve up to 2x speedup in parallel inference. The same source reports that DwarfStar 4 can run DeepSeek V4 Flash on 128GB MacBooks at practical speeds. It remains unconfirmed whether the GeForce RTX 5090 can run Gemma 4 with MTP using INT4 quantization, as the author did not test that configuration.