Releases·Tool Release
Intel releases OpenVINO 2026.4 with new GenAI models, NPU support, and speculative decoding
Intel shipped OpenVINO 2026.4, expanding Gen AI model coverage and adding inference performance features across CPU, GPU, and NPU. Highlights include Multi-Token Prediction speculative decoding for Gemma 4 and Qwen 3.5/3.6, Tree Drafting for EAGLE3 VLM pipelines, NPU profiling via Intel VTune, and new Node.js ASR pipeline support. The release targets developers running AI at the edge, in the cloud, or locally with minimal code changes.
4 engagement1 sourceThu, Sep 17, 2026, 09:51 AM
AI Overview
- OpenVINO 2026.4 adds model support across CPU, GPU, and NPU: Gemma-3n, Kokoro-82M, Qwen3-VL-4B with EAGLE3, Muse Glimmer 30B, Gemma4 12B, DeepSeek OCR-2, FLUX.2-Klein 4B, and Granite 4.0 H Micro.
- OpenVINO GenAI adds Multi-Token Prediction speculative decoding for Gemma 4 and Qwen 3.5/3.6 on CPUs and GPUs, raising throughput without accuracy loss.
- Tree Drafting (Top-K) for EAGLE3 unlocks higher VLM pipeline throughput than Chain Drafting (Top-1).
- Xe3 integrated graphics optimizations improve long-context Gemma 4 inference on Intel Core Ultra Series 3 processors.
- ITT profiling now extends to the NPU, letting Intel VTune Profiler analyze CPU, GPU, and NPU execution in one toolchain.
- Previews include DFlash acceleration for Qwen models on GPUs, bounded dynamic-shape NPU vision workloads (validated with ESPCN), and idle model unloading in Model Server.
- ASRPipeline support in Node.js lets JavaScript developers run streaming Whisper and Qwen3-ASR with a pipeline API matching C++ and Python.
- Early-release models include Qwen-image, Z-Image-Turbo, LFM2.5-8B-A1B, MiniCPM5-2B, RF-DETR, and BGE M3 Reranker.
Related
Releases·Tool Release
Intel ships Day-0 OpenVINO support for Qwen-Image-2.1, one open-weight checkpoint for generation and editing
Intel announced Day-0 OpenVINO optimization for Qwen-Image-2.1, enabling the open-weight image model to run optimized on Intel hardware. The single checkpoint supports both image generation and editing, so developers can deploy one artifact for both tasks.
Intel announced Day-0 OpenVINO optimization for Qwen-Image-2.1, enabling the open-weight image model to run optimized on Intel hardware. The single checkpoint supports both image generation and editing, so developers can deploy one artifact for both tasks.
108 engagement1 sourcexTue, Sep 22, 2026, 04:08 AM
Releases·Tool Release
qMLX: Custom inference engine for Qwen 3.5 122B on Apple Silicon released
A developer released qMLX, a custom inference engine for Qwen 3.5 122B on Apple Silicon, extending MLX with hybrid attention support, SSD-backed KV cache, and RYS layer duplication. It enables efficient serving of the hybrid MoE model on consumer Mac hardware.
A developer released qMLX, a custom inference engine for Qwen 3.5 122B on Apple Silicon, extending MLX with hybrid attention support, SSD-backed KV cache, and RYS layer duplication. It enables efficient serving of the hybrid MoE model on consumer Mac hardware.
19 engagement1 sourcegithubThu, Jul 9, 2026, 09:47 PM
Social·Community
ZGX Nano G1N AI Station - Which model for company testing
A small-business user asks which local model to run on a ZGX Nano G1N AI Station for document processing, general Q&A, and coding with a few users. They currently run Qwen 3.6 35B A3B nvfp4 and want something more precise, considering Gemma 4 or Qwen 3.8 27B. They also ask whether RAG needs separate setup in Open WebUI, with MCP and SearXNG web search already configured.
A small-business user asks which local model to run on a ZGX Nano G1N AI Station for document processing, general Q&A, and coding with a few users. They currently run Qwen 3.6 35B A3B nvfp4 and want something more precise, considering Gemma 4 or Qwen 3.8 27B. They also ask whether RAG needs separate setup in Open WebUI, with MCP and SearXNG web search already configured.
5 engagement1 sourceredditTue, Sep 15, 2026, 09:44 PM
Releases·Tool Release
vLLM v0.29.0: Model Runner V2 default, Kimi K3 and DeepSeek V4 performance, Mamba prefix caching
The vLLM project shipped v0.29.0 on GitHub, making Model Runner V2 the default for all models after rollout began with pooling models. The release includes 594 commits from 277 contributors, 91 of them new. It adds CUDA graph memory profiling for KV cache auto-sizing and batch-sharded sampling that cuts per-step logits memory by 1/TP.
The vLLM project shipped v0.29.0 on GitHub, making Model Runner V2 the default for all models after rollout began with pooling models. The release includes 594 commits from 277 contributors, 91 of them new. It adds CUDA graph memory profiling for KV cache auto-sizing and batch-sharded sampling that cuts per-step logits memory by 1/TP.
3 engagement1 sourcegithubWed, Sep 9, 2026, 08:54 AM
Releases·Tool Release
vLLM v0.27.0 adds Kimi K3 support with full-stack integration
vLLM v0.27.0 ships with 561 commits from 242 contributors, adding full-stack support for Kimi K3 including kernels, frontends, and quantization. It also introduces Qwen3.5 models and K-EXAONE-2.0-750B, expanding the inference engine's model coverage.
vLLM v0.27.0 ships with 561 commits from 242 contributors, adding full-stack support for Kimi K3 including kernels, frontends, and quantization. It also introduces Qwen3.5 models and K-EXAONE-2.0-750B, expanding the inference engine's model coverage.
8 engagement1 sourcegithubMon, Aug 10, 2026, 09:18 PM