New models, hardware, research, and techniques for running AI locally. We are selective: a quieter feed of things that change what is possible, practical, or legal is more useful than a complete one.
DeepSeek has published V4.1 Flash under the MIT licence: a 552-billion-parameter multimodal mixture-of-experts that reads images and text and activates only 8 to 16 billion parameters per token. It uses an encoder-decoder design rather than the decoder-only shape almost every other current model uses, and it carries a 1-million-token context. It is server-class, but genuinely open.
FeaturedMiniMax H3, a widely shared open-weight video model, is licensed for use everywhere except the EU, the UK, Korea and the US, and the restriction covers the model's outputs as well as its weights. It is not alone. For a UK reader, this is a case where the licence, not the hardware, is what stops you, and it is invisible to a licence badge.
At IFA 2026 NVIDIA announced local-AI updates, and the quietly important part is software: kernel optimisations it has contributed to the open-source llama.cpp and vLLM projects, which it says lift throughput by up to 1.9x on a GeForce RTX 5090. Because they are upstream, existing NVIDIA GPU owners get the gain by updating, with no new hardware. A new RTX Spark desktop with 128GB of unified memory was also shown.
Z.ai has published the open weights of GLM-5.3, roughly two weeks after its hosted launch and following a security review. The full model is a server-class mixture-of-experts of around 744 billion parameters under a bespoke GLM-5.3 licence, not the MIT terms its predecessor carried. A smaller, natively multimodal GLM-5.3-Flash also shipped, and that one is MIT.
FeaturedApple has announced M5-generation Mac Studio and Mac mini machines and is positioning them squarely at local AI. The headline for people running large models is the top Mac Studio: an M5 Ultra with up to 512GB of unified memory at a claimed 1.2TB/s, enough to hold very large open-weight models in memory on a single desktop. Pre-orders opened on 25 August, with machines shipping 22 September.
FeaturedGenerating video and music locally used to mean data-centre hardware or a non-commercial licence. That is changing: there are now open video models that fit a single 24GB card and permissively-licensed music models that run in 8GB. A short survey of where local generative media stands, and the licence traps to watch.
Zhipu has released GLM-5.3 through its hosted coding service, claiming the strongest open-weight coding model available. Two caveats matter for a local audience: the weights are not downloadable yet, said to follow in roughly two weeks after a security review, and every benchmark figure so far is the company's own. We will cover it properly, and add it to the catalogue, when the weights are actually released.
The Qwen3.8 open weights our earlier report treated as pending have now landed, and they split cleanly in two. Qwen3.8-27B is a permissively licensed, multimodal model that runs on a single 24GB card. The 2.4-trillion-parameter flagship is a text-only, cluster-scale release under a custom licence. For most people running local AI, the 27B is the story.
FeaturedLiquid AI has released LFM2.5-VL-3B, a 3.1-billion-parameter open-weight vision-language model designed to run on phones, laptops, and single GPUs rather than in a data centre. It ships with day-one GGUF, MLX, and ONNX exports and support across the common local runtimes, and adds screen understanding, object grounding, and function calling. The licence is open but carries a revenue condition on commercial use.
NVIDIA has released Nemotron 3.5 Lightning, an open 30-billion-parameter mixture-of-experts with about 3.6 billion active per token, a hybrid Mamba-Transformer design, and a one-million-token context. It is unusually open, shipping under a permissive licence with the training data and recipes alongside the weights, and it is aimed at fast, high-volume agentic work. Runtime support for its hybrid architecture is still maturing in local engines.
Meta has released Muse Glimmer-30B, a multimodal model distilled from its larger closed Muse Spark and aimed at agentic use on a single consumer GPU. The headline for local AI is the licence: it ships under Apache 2.0, dropping the conditions that Meta's Llama models carry.
Ollama v0.32.6 has the MLX engine use a model's own multi-token-prediction head to drive speculative decoding automatically, which speeds up Qwen on Apple GPUs. It also aligns streaming with OpenAI's format and widens GPU support. An incremental release, with a couple of genuinely useful improvements.
Alibaba has announced Qwen3.8-Max, a 2.4-trillion-parameter mixture-of-experts model, with open weights said to be arriving on Hugging Face, alongside a much smaller Qwen3.8-27B for ordinary hardware. As of writing the weights are not yet published and no licence has been named, so treat both as announced rather than available.
Moonshot AI has released the open weights for Kimi K3, a 2.8-trillion-parameter mixture-of-experts model that tops several open-weight leaderboards. It is a clear case of self-hosting at server scale: running it needs a cluster, not a workstation, and its licence terms were not clearly published at launch.
Mistral has released the Mistral 3 family, spanning small dense models and a large mixture-of-experts, all under the Apache 2.0 licence. The licensing is as notable as the models: it removes the commercial conditions of Mistral's earlier releases.
Mistral has released Leanstral 1.5, an Apache 2.0 model specialised for Lean 4, the proof assistant used in formal software verification. It is a reminder that local models are not only general chatbots, but increasingly tools for specific domains.
The Council of the EU has given final approval to the Digital Omnibus, a package that simplifies and defers parts of the AI Act. The main high-risk obligations move from August 2026 to December 2027, while the prohibitions and model rules stay live.
FeaturedA tightening memory market has led Apple to narrow the memory configurations on its top Mac Studio, removing some of the very high-capacity options that made it a favourite for running large models. If you were counting on a specific config, check current availability.
DeepSeek has released its V4 generation, reportedly as two models, V4-Pro and V4-Flash, under the permissive MIT licence and as open weights. The company frames it as an architectural step rather than a scale-up. Treat the performance claims as claims until independent results appear.
Alibaba has followed Qwen3.5 with Qwen3.6, reported to include new open-weight checkpoints, among them a mixture-of-experts model. The rapid cadence continues to be as notable as any single release in the line.
Mini PCs built on AMD's Ryzen AI Max+ 395 pair a large integrated GPU with up to 128GB of unified memory, most of it usable as VRAM. They run models that no single consumer discrete card can hold, at low power, though bandwidth caps their speed.
Alibaba's Qwen team has released the Qwen3.5 series, continuing the family's fast release cadence and its Apache 2.0 licensing. As always, treat the accompanying benchmark figures as the developer's own claims until independent results appear.
Black Forest Labs has released FLUX.2 [klein], a family of compact image models, reported at 4B and 9B sizes, built for fast, near-real-time generation on consumer GPUs. It is the openly licensed member of the new FLUX.2 line.
NVIDIA's DGX Spark, a compact desktop built on the GB10 Grace Blackwell chip with 128GB of unified memory, has gone on sale through NVIDIA and partners. It targets developers who want NVIDIA's software stack on a local machine, with modest bandwidth as the main trade-off.
Meta has released Llama 3.3 70B Instruct, an open-weight model that the company says approaches the quality of its 405B model while needing far less hardware to run. It ships with a 128k context window under the Llama 3.3 Community License.
Showing all articles. As the archive grows this listing will paginate.