The Ollama v0.40.6 pre-release includes updates to the GGUF model format and Windows runner stability. It adds support for q2_0 tensors and introduces identifier-based question keys for the --ask mode.

The vLLM project has released version 0.32.0, which specifically addresses a bug in the Mamba2 chunk scan kernel. This bug caused the kernel to produce NaN (Not a Number) outputs when running on sm90 architecture. The fix ensures numerical stability for Mamba2 models on this specific hardware.

The v0.40.6-rc1 release candidate for Ollama introduces changes to how the runtime handles registry redirects when a proxy is present. It also addresses a stability issue by fixing a nil CancelFunc crash during blob download releases.

The latest Ollama release adds support for NVIDIA's Nemotron 3.5 Super and Nemotron 3.5 Bolt vision-language models. It also introduces the ability to run decision models through the ollama.com cloud using specific identifiers, and improves error reporting for cloud access and OpenAI-compatible tool calls.

Ollama has released version 0.40.5-rc1, which includes support for the Nemotron 3.5 model. The update focuses on improving the MLX backend, specifically addressing Nemotron bugs, stabilizing Mamba components, and enhancing global scale support for various quantization formats.

Ollama version 0.40.4 introduces the ability to run decision models through the terminal for yes/no questions, text classification, and rubric scoring. The update also brings significant performance and reliability improvements to the MLX runner on Apple silicon and enhances Gemma 4 prefill performance on NVIDIA GPUs.

The v0.40.4-rc1 release candidate for Ollama introduces Kev decision scoring capabilities specifically for MLX. This update allows for loading Kev models with FP32 decision heads and release metadata. Additionally, it implements System One scoring support featuring shared-state caching and specific limits for Kev.

The Jan open source desktop AI assistant has released version 0.8.6 as a hotfix for version 0.8.5. This update resolves model loading crashes in the Linux AppImage by using the system's Vulkan loader instead of a bundled one. It also fixes compatibility issues for NVIDIA Turing GPUs (GTX 16xx, RTX 20xx) on older drivers and improves memory management for Qwen3.8 with speculative decoding.

The latest release of Jan Open Source Desktop AI Assistant, version 0.8.5, focuses on fixes and improvements rather than new features. The llama.cpp engine is now bundled with the app, and several issues related to local models, GPU support, and security have been addressed. Additionally, the desktop executable has been renamed to Jan-Desktop, and the jan CLI is now a separate install.

The vLLM 0.31.0 release includes 717 commits and introduces deep optimizations for DeepSeek-V4.1-Flash, including FlashMLA mega attention with NVFP4 compressed KV cache. The update also features Model Runner V2 with improved speculative decoding, new large-scale serving backends like MoonEP, and a new preload CLI to keep weights resident in GPU memory across restarts.

The LocalAI 4.11.0 release introduces significant updates for audio intelligence, allowing for combined transcription, speaker diarization, and sound detection. It also implements model failover chains to ensure availability by switching between local and remote targets. Additionally, the update adds signed OCI model galleries, Kimodo text-to-animation capabilities, and a new 'Decisions' API for structured model responses.