News

FLUX 3 Image Klein is a 5.6B open-weights model for text-to-image generation and editing: one model for prompting, single-reference edits and multi-reference composition, with placement boxes in the prompt. Ships in BF16, FP8 and NVFP4 (5.3GB).

inclusionAI releases Ling 3.1 Flash: 560B total parameters, 25B active per Token, hybrid MoE architecture, up to a 1M-Token Context Window and an MIT license. At 4 bits it fits in 512GB of Unified memory — the edge case of today's Mac Studio M5 Ultra article.

NVIDIA has published MagpieTTS Multilingual on Hugging Face, a compact 357M-Parameter speech synthesis model covering 12 languages, including Italian and French. Encoder-decoder architecture, designed for low-Latency local execution.

ARDY is a 326M-Parameter autoregressive diffusion model for interactive real-time humanoid motion generation: online text prompting, flexible kinematic constraints (root paths, keyframes, sparse poses) and sequences up to 8 seconds long. The Core-RP-20FPS-Horizon40 checkpoint, trained on the Bones Rigplay 1 dataset, is on Hugging Face since October 7 and optimized for NVIDIA Ampere, Hopper and Blackwell GPUs.

NVIDIA publishes Kumo-Tabular on Hugging Face, a 500-million-Parameter foundation model for tabular data with OpenMDW 1.1 support and ready-to-use checkpoints: an example of local AI on structured business data.

OpenBMB releases MiniCPM-V 4.7, 35B-A3B variant: the first open-source MoE able to Decode smoothly on a smartphone without a quality compromise — 13+ tokens/second with only 3GB of memory on an iPhone 17 Pro Max. The secret: Lookahead Expert Scheduling on llama.cpp-omni, combined with fine-grained layer-wise Quantization and speculative sampling. Day-0 support for Rockchip, Intel, AMD and Arm; Ollama, llama.cpp, vLLM and SGLang.

Autotrust releases JEV-35B, a 35B-Parameter (3B active) MoE Decision model fine-tuned on Qwen3.5-35B-A3B for text classification and risk scoring. In NVFP4 Quantization it occupies about 25.6GB of VRAM: it runs on a 32GB GPU without going through the frontier.

Aleph Alpha's new sovereign Open-weight model, trained from scratch in Germany: 78B parameters (3.46B active) with up to 1M tokens of context. Built for mission-critical environments - multi-step reasoning, structured extraction, RAG and reliable tool calling, with native German and English. Weights are on Hugging Face, with a full technical report.

DeepSeek V4.1 Flash pushes the MoE architecture further: 552B total parameters but only 8-16B active per Token, a 1M-Token Context Window, and an MIT license. A model built for online serving, but useful as a reference for learning to read model spec sheets.

Gemma 4, released in March 2026, marks a license change for Google's family: for the first time the weights are under Apache 2.0. Available in E2B, E4B, 26B A4B (MoE) and 31B sizes, with context windows up to 256K tokens and text, image, and audio input. The 12B Unified followed in June — an ideal size for PCs with consumer GPUs.

Prism ML releases Bonsai-2 27B in GGUF: a 27B reasoning model with Ternary weights at 1.72 bits per weight, 5.95 GB and a 262K-Token context, retaining 98.2% of the FP16 intelligence and reaching ~47 tok/s on an Apple M5 Max laptop.

A new open speech-to-text model: real-time streaming with low Latency, built for live dictation and captions. One more step toward audio that genuinely runs locally, without going through the cloud.

Qwen-Image 2.1 is a 7B Open-weight model (32-layer DiT) that unifies text-to-image generation up to 2048x2048, editing with up to 10 reference images, and native RGBA transparency. With ~14GB of VRAM in bfloat16 it runs on consumer hardware, and ComfyUI supports it natively: a concrete step for cloud-free image production.

Alibaba's Qwen team released Qwen3.8-Omni-Flash, a native omnimodal model that jointly handles text, images, audio and video with a one-million-Token Context Window. It is built around agentic capabilities: it plans tasks, calls tools and produces finished audio-visual content. The Qwen family keeps growing right as this guide analyzes it for local use.

Meta released Llama 4, the new Open Source language model with 100 billion parameters. Performance surpasses GPT-4 on many benchmarks and the model is optimized for local execution with Quantization.