Toolkit

It is Halloween, so this one is seasonal and entirely legitimate. NVIDIA's Nemotron team built MonsterHouse: an interactive animated garage door that listens to trick-or-treaters, responds in character as a grumpy pirate ship reincarnated as real estate, and softens if visitors say please. The full build guide is on Hugging Face.

What's worth pulling apart here is the stack. NVIDIA's Nemotron 3.5 Bolt E4B 4GB is a small open-weights model (one you can download and run yourself) designed for conversational agents. It runs the dialogue. A LoRA adapter (a lightweight fine-tuning layer applied on top of a base model to change its personality or task) supplies the pirate house character. Nemotron ASR handles speech-to-text; Magpie TTS handles text-to-speech. The whole pipeline runs on a single workstation.

If you want to try the conversational model on its own, Nemotron 3.5 Bolt E4B 4GB is available on Hugging Face now, at no cost.

Worth knowing: The full MonsterHouse setup requires an NVIDIA RTX PRO 6000 GPU — hardware most people don't own. The language model alone runs on more modest hardware, but the animated talking door is firmly in enthusiast territory.

Mistral Large 4 (nicknamed "Le Chonk") is a large open-weights model — meaning you can download and run it yourself — from French AI company Mistral. It is multimodal (accepts both text and images as input) and handles multiple languages. The context window (the amount of text it can read and respond to in one go) is one million tokens, roughly equivalent to several full-length novels. It uses a Mixture-of-Experts (MoE) architecture, where only a subset of the model activates for any given input, making it more efficient than its total size suggests.

The standout capability is cybersecurity. Mistral Large 4 scores 82% on the Artificial Analysis Cyber Index — a test requiring a model to reproduce a real software vulnerability and then patch it — according to Mistral's own testing. Several closed commercial models score near zero on the same benchmark because they refuse to attempt the task at all. For defensive security work, that refusal is a problem: patching a vulnerability usually requires understanding how to exploit it first.

Quantized variants (compressed versions that run on consumer hardware) are available via Hugging Face. Download it here.

Worth knowing: The licence is an MIT derivative with restrictions on using the model as the core of a commercial API service — check the terms before building a product on top of it.

OpenAI has released GPT-6.1 Terra, a model built for computer-use tasks: filling in forms, reading invoices, editing video, interpreting technical drawings.

The headline improvement is factuality. On OpenAI's own internal benchmark of error-inducing prompts, the error rate fell from 53.6% with the previous GPT-5.6 Terra to 16.7% — a meaningful drop, though these are OpenAI's own evaluations, not independent tests.

API pricing is 50% lower than GPT-5.6 Terra. For everyday users, it is available now through ChatGPT.

To try it: open ChatGPT, select GPT-6.1 Terra from the model picker, and give it a task involving a document or image — uploading a PDF and asking it to summarise or reformat works well as a first test.

Worth knowing: The computer-use benchmark scores — including comparisons to Claude Opus 5.5 and GPT-6 Astra — come from OpenAI's own testing. No independent evaluation of Terra has been published yet.

Kimi K3.1 Pro is a large language model — an AI that takes text input and produces text output — from Chinese lab Moonshot AI, now available through their API. It is not yet open-weights (meaning you cannot download and run the model yourself), but Moonshot has committed to releasing the full weights by November 4th.

Independent benchmarking firm Artificial Analysis scored it at 55 on their quality index, placing it ahead of Claude Fable 5.1, GPT-6 Astra, and Gemini 4 Argon (all at 53), and behind only Claude Opus 5.5 (58). This is the highest score an open-weights-bound model has reached against frontier competition.

Pricing: $2.00 per million input tokens and $10.00 per million output tokens — the same as GPT-6.1 Sol and half the cost of Claude Opus 5.5. A Highspeed version runs 2–3x faster at twice the price.

Access it at moonshot.ai or via any OpenAI-compatible API wrapper.

Worth knowing: The model weights are not yet public — if you need to run this locally or offline, wait until after November 4th.

OlmoEarth v1.3 is an open-weights satellite image model (the model weights — the core learned parameters — are publicly downloadable) from the Allen Institute for AI. The new Embeddings Explorer is a browser-based tool built on top of it that lets you explore precomputed maps of the entire planet based on the model's analysis of satellite data.

No account, no sign-in, no GPU required. The model has already processed a year's worth of radar and multispectral satellite imagery for every 10-metre patch of land on Earth. You browse the results directly in your browser.

Practical uses include tracking land cover changes over time, spotting deforestation patterns, or comparing regions by terrain type. The underlying model weights, training data, and code are fully public.

To start: visit the Embeddings Explorer linked from the HuggingFace post — no setup needed for the browser tool.

Worth knowing: The Explorer is a visualisation and exploration tool, not an analysis platform — if you want to run your own classifications or produce maps programmatically, that requires downloading the model weights and some coding ability.

TorchNeuron is a new PyTorch plugin (PyTorch is a widely used framework for building and running AI models) that lets you run and train models on AWS Trainium — Amazon's own AI chips designed for cost-effective training — using the same code you would write for any other hardware. Previously, running models on Trainium required working through a separate library with significant workarounds; TorchNeuron makes Trainium behave like a standard device, so standard Hugging Face tools work on it without modification.

Available now in public beta. Install with:

pip install torch-neuronx

You will need access to an AWS Trainium instance — this is not a laptop tool. Costs depend on your AWS usage. 89% of the top 100 most-used text-generation model architectures are supported, according to Hugging Face's own figures.

Full details and code examples at the Hugging Face blog.

Worth knowing: This requires AWS Trainium hardware — there is no way to use TorchNeuron on a standard GPU or consumer machine.

Beam, built by Reflection and released on Hugging Face, is a large open-weights model — meaning the trained model files are publicly available to download and run — focused on coding and agentic work (tasks where an AI takes a sequence of actions toward a goal).

Its headline spec is a 1-million-token context window (the amount of text a model can read and reason over in one session), which is large enough to hold an entire codebase or a book-length document. It uses a sparse MoE (mixture-of-experts) architecture, where only a fraction of the model activates for each piece of text, so it runs faster than its full size suggests.

The licence is Apache 2.0, meaning commercial use is permitted. Get it at Hugging Face.

To run the full model at usable speed, you will need significant server-grade hardware — this is not a laptop model.

Worth knowing: The full 501-billion-parameter version requires substantial GPU infrastructure; smaller quantized versions (compressed variants that trade some precision for lower memory use) are available but check your hardware specifications before committing.

FLUX.3 Image Klein, from Black Forest Labs, is a free open-weights image generation model now available on Hugging Face. It handles text-to-image, image-to-image editing with reference photos, and layout control — letting you specify exactly where elements appear in a composition using a simple coordinate system.

The easiest way to try it is directly in the Hugging Face Space, no installation needed. For those comfortable running models locally, it works with the diffusers Python library and supports smaller GPU configurations via FP8 and NVFP4 quantized (compressed) weights.

What sets it apart from standard generators: you can pass multiple reference images and instruct the model to combine elements from each, or lock specific objects to specific areas of the frame — useful for anything requiring precise composition.

Free to use. Local installation needs a capable GPU; the browser demo works without one.

Worth knowing: The licence terms differ between the two available checkpoints — check which applies before using outputs commercially.

LFM2.5-24B-A2B is a free, open-weights model from Liquid AI — you can download and run it today on a capable laptop or desktop, with no cloud account required and no data leaving your machine.

It is designed for agentic tasks: give it a goal, and it reasons step by step before acting, calling tools along the way. On an Apple M5 Max, it runs at 184 tokens per second — fast enough to feel responsive.

To start, install Python and run:

pip install -U transformers

Then load the model using the code in the Hugging Face model card. It is compatible with llama.cpp and MLX, which offer simpler setups if you prefer not to use Python directly.

Worth knowing: You will need roughly 16GB of RAM to run this at usable speed — check your machine's specs before downloading.

ttok is a command-line tool that counts how many tokens (the chunks a language model splits text into — roughly three-quarters of a word each, on average) a piece of text contains. It is useful for checking whether a document fits inside a model's context window (the maximum amount of text a model can process in one go) before you try to use it.

Version 1.0 was released this week by Simon Willison, primarily to switch the default tokenizer to GPT-5 rather than GPT-4. The tool is available via simonwillison.net and installs in one command: uv tool install ttok. Pipe any text into it and it returns a token count.

It is free, runs locally, and requires no account.

Worth knowing: ttok defaults to the GPT-5 tokenizer; if you are working with a different model family, counts may be approximate — and OpenAI has not yet officially confirmed whether GPT-6 uses the same tokenizer as GPT-5, so even that default carries a small caveat.