June 2026 Summaries

Gemma 4 in Ollama 0.31 demonstrates a significant speed improvement on Apple Silicon, generating tokens nearly 90% faster due to multi-token prediction (MTP). This enhancement is achieved by employing a small, fast draft model that proposes several tokens, which are then verified by the main model in a single pass, significantly boosting efficiency without altering outputs. The draft model's proposals, often accepted in predictable coding contexts, make coding agents more responsive as they continuously interact with files and tools. The efficiency of MTP is automatically fine-tuned during runtime to optimize performance based on workload, ensuring speculative decoding does not slow down the process when it stops being advantageous. This approach leverages GPU capabilities for simultaneous drafting, sampling, and verification, with a new matrix multiplication kernel enhancing the speed of Gemma 4's computations on specific setups. This advancement benefits real programming tasks, with potential application in other models beyond Gemma 4.

Ollama's updated MLX engine, designed for optimal performance on Apple Silicon, leverages Apple's unified memory and the Metal-backed MLX framework to produce faster, higher-quality responses while using less memory. This update includes support for NVIDIA’s model-optimized NVFP4 format, which enhances output quality compared to other 4-bit quantization formats by closely tracking model weights' local dynamic range and reducing quantization loss. The engine is now 20% faster due to optimizations like fusing operations into single Metal kernels and more efficient GPU-backed sampling. Additionally, Ollama's new snapshot system improves responsiveness in agent workflows by saving model state at key points, allowing for seamless resumption and efficient handling of multi-agent sessions, reasoning models, and conversation branching. This system ensures that only new directions in conversations need processing, conserving memory and enhancing performance. Users can access these improvements by downloading the latest version of Ollama and running models such as Gemma 4 12B with the MLX engine.

Ollama 0.30 introduces enhanced performance and broadened compatibility for machine learning models on various hardware platforms, thanks to contributions from NVIDIA and llama.cpp. This update enables faster processing on NVIDIA GPUs, achieving up to a 20% performance boost, and expands GPU acceleration to AMD and Intel devices via Vulkan, eliminating the need for additional vendor-specific libraries. The release also enhances compatibility with the GGUF model ecosystem, allowing users to run diverse models, including LFM, Prism, and those from Unsloth, directly from platforms like Hugging Face. Users can easily integrate these models into coding agents and personal assistants, with tool-calling capabilities verified through the Ollama show command. Ollama acknowledges the collaborative efforts of Georgi Gerganov, llama.cpp, and hardware partners such as NVIDIA, AMD, Qualcomm, and Intel in optimizing performance across platforms.

NVIDIA Nemotron 3 Ultra, a 550 billion parameter model with 55 billion active parameters, is now accessible on Ollama's cloud platform, offering a solution optimized for long-running and agent-driven workflows. This model is designed to handle complex tasks across multiple steps efficiently, with a 1 million token context that maintains extensive codebases and tool histories. Utilizing NVIDIA's NVFP4 format, it achieves high efficiency by operating within less memory and delivering faster performance. Nemotron 3 Ultra distinguishes itself in benchmarks by leading in accuracy and throughput for tasks related to agent productivity, coding, and instruction adherence, while also being cost-effective, saving up to 30% compared to other open models.