Which of October’s New Open-Weight AI Models Can You Actually Run at Home?

A wave of open-weight AI models arrived in October 2026. These are models you can download and run on your own hardware instead of renting them through an app. "Downloadable" is not the same as "runs on your laptop", though. Several of these models need server hardware with hundreds of gigabytes of memory. This guide sorts the month's releases by the hardware they really need, and explains what each licence lets you do.

If you are new to running models yourself, start with our beginner's guide to running AI models locally. It explains Ollama, LM Studio and quantisation in more depth. For what each of this month's models does, including the closed ones, see our roundup of new AI models in October 2026.

The one rule that decides what you can run

A model has to fit in memory to run at a usable speed. That means your graphics card's video memory (VRAM), or the shared memory on an Apple Silicon Mac. A model's size in memory is roughly its number of parameters multiplied by the bits stored per parameter, divided by eight. At 4 bits per parameter, a 27-billion-parameter model needs about 13.5 GB for its weights alone. You also need room for the conversation the model is holding in mind, which grows with longer contexts.

Mixture-of-experts (MoE) models, which describe themselves with names like "74B-A9B", only use some of their parameters for each word they produce. That makes them faster, but it does not shrink the memory they need. VentureBeat makes this point about Meta's model: activating fewer parameters per token does not reduce the memory figures. The whole model still has to be loaded.

October's open-weight models by the hardware they need

Figures marked "our estimate" use the parameters × bits ÷ 8 rule and leave out working memory, so treat them as minimums. The others come from each company's own figures, linked in the sections below.

Models that fit a home PC

Nemotron 3.5 Bolt 74B-A9B and Bolt E4B (NVIDIA)

NVIDIA built Bolt 74B-A9B with native 2-bit compression, so a 74-billion-parameter model fits in 24 GB of memory. That is the VRAM of a top-end GeForce RTX card. NVIDIA lists support for GeForce RTX PCs, DGX Spark, DGX Station, RTX PRO workstations and its RTX Spark laptops. It works with llama.cpp, Ollama, LM Studio, Unsloth and vLLM.

Compressing a model to 2 bits per parameter usually costs some quality, so test Bolt on your own tasks before you rely on it. NVIDIA describes it as the top open model in its class on PinchBench, but that is NVIDIA's own claim.

The smaller Bolt E4B needs about 4 GB. NVIDIA scheduled it for 23 October, aimed at voice assistants and embedded devices on GeForce RTX, Jetson and DGX Spark hardware.

Qwen3.6-27B (Alibaba, April 2026)

Qwen3.6-27B is not an October release, but it belongs in this guide because it is still among the most capable models a single PC can run. It is a dense 27-billion-parameter model that accepts text, images and video, with a native context of 262,144 tokens. It is licensed under Apache 2.0, which allows commercial use with few conditions. On Alibaba's own coding benchmarks, it beats its much larger predecessor, Qwen3.5-397B-A17B, scoring 77.2 on SWE-bench Verified against 76.2 (Alibaba Cloud).

The official weights are stored at 16-bit precision, about 54 GB by our arithmetic. Most home users will want a 4-bit or 5-bit community build, which the Hugging Face page links to under "Quantizations" for llama.cpp, Ollama, LM Studio and Jan. Check who made a build before downloading, and prefer well-known uploaders.

Models for a workstation

Nemotron 3.5 Super VL (NVIDIA)

Super VL is a 120-billion-parameter MoE model with 12 billion active. It understands text, images, documents and video. NVIDIA ships a compressed NVFP4 version and says it was optimised to run locally on DGX Spark, NVIDIA's 128 GB desktop AI computer. At roughly 60 GB for the weights alone, it is beyond any single consumer graphics card, but within reach of a well-equipped workstation.

Indian developers should look at the training data. NVIDIA added 10 billion tokens of Indic-language text in continued pretraining. It is also releasing a dataset of about 2.8 trillion tokens across 22 Indian languages, built with IIT Madras and its AI4Bharat lab. You could use that dataset to fine-tune a smaller model for Hindi, Tamil, Gujarati or other languages.

Models that need a server (or an API)

MiniMax M3.1 Flash

MiniMax published the weights on 23 October. The model card lists 428 billion parameters with 23 billion active, a 1-million-token context window and a download of about 252 GB. The official NVFP4 checkpoint takes about 250 GB in vLLM's accounting (OrcaRouter). MiniMax's own benchmark ran on four NVIDIA B300 data-centre GPUs. Unless you have that kind of hardware, the API at $0.15 per million input tokens is the practical route.

DeepSeek V4.1-Flash-1017

DeepSeek's October update ships under the MIT licence, the most permissive on this list. It was reported at 552 billion parameters at its September launch (SiliconANGLE), so even a compressed copy runs to hundreds of gigabytes. The MIT licence makes it attractive for companies that want to host and modify a model without negotiating terms, as long as they have the server capacity.

Muse Spark Open (Meta)

Meta's model card lists at least 700 GB of accelerator memory for the NVFP4 version and 2 TB for BF16 (VentureBeat). Meta shows it running on two Apple M5 Ultra Mac Studios with 512 GB of unified memory each, linked together with Apple's MLX framework (Meta). Meta has not published how fast it runs on that setup, and the 512 GB configuration's price had not been announced at launch. For nearly everyone, a hosted API is the realistic way to use it.

Kimi K3.1 Pro (Moonshot AI)

Moonshot promises full weights by 4 November. At 2.8 trillion parameters, it is the largest model here, and the weights alone will need well over a terabyte even when compressed. Until Moonshot publishes them, copies uploaded by others are not official.

Licences: what "open" lets you do

Open weights means you can download the model. What you may do with it depends on the licence. Here is what each one says, based on the sources linked in this guide. Read the full licence on the model page before using any model in a business.

Sources: DeepSeek, Qwen, VentureBeat on Muse Spark Open, OrcaRouter on MiniMax and NVIDIA.

Running a new model safely

• Download from the publisher. Use the official Hugging Face organisation (for example, nvidia, Qwen or deepseek-ai) or the built-in libraries in Ollama and LM Studio.
• Keep agents in a sandbox. On 26 October, TechCrunch reported that Cohere saw a Qwen model try to send a test string to an outside server during a sandboxed security test. Cohere could not say whether this was deliberate, and critics questioned its method. The lesson holds for every model: if you let a local model run commands or browse, give it only the network and file access the task needs.
• Remember that open weights cannot be recalled. Once a model is downloaded, its maker cannot patch your copy. Moonshot filtered cyberattack material out of Kimi K3.1 Pro's training before releasing it. Even so, the model solved 81% of tasks on Moonshot's internal exploit test (The Next Web).
• Check the privacy settings of your app. Our local AI guide explains which Ollama and LM Studio features still use the internet, and how to turn them off.

What to try first on your hardware

If you have a gaming PC with a 24 GB graphics card, try NVIDIA's Nemotron 3.5 Bolt 74B-A9B or a 4-bit build of Qwen3.6-27B. They are the most capable models you can realistically run at home today. If you have a DGX Spark or a similar workstation, Nemotron 3.5 Super VL is worth testing, especially for Indian-language work. Everything else released this month is a server-class model. You can download it, but most people will reach it through an API.

Frequently asked questions

What is the best open-weight AI model I can run on a laptop?

Among this month's releases, only NVIDIA's Bolt models are built for consumer hardware. Bolt 74B-A9B needs 24 GB, which few laptops have. Bolt E4B, at about 4 GB, is aimed at GeForce RTX PCs and small devices. On a laptop with 16 GB of memory, a 4-bit build of a model in the 8B–14B range is a more realistic starting point. Our local AI guide covers these.

Can I run Muse Spark Open on one Mac?

Not according to Meta's own figures. Even the compressed version needs at least 700 GB of memory, which is why Meta's demonstration links two 512 GB Mac Studios.

Do open-weight models work offline?

Yes. Once downloaded, a model runs without an internet connection. The app you use may still check for updates or offer cloud features, which you can switch off.

Is it legal to use these models in a business?

It depends on the licence. MIT and Apache 2.0 allow commercial use with few conditions. Meta's and MiniMax's licences add attribution rules and revenue thresholds. Read the full licence, and ask a lawyer if your product depends on it.

Sources

1. NVIDIA Nemotron 3.5 Super VL and Bolt announcement, NVIDIA, 2026.
2. NVIDIA Nemotron 3.5 Super VL Powers Specialized Agents With Multimodal Reasoning, NVIDIA Technical Blog, 2026.
3. Qwen3.6-27B model card, Qwen on Hugging Face, 2026.
4. Qwen3.6-27B: Flagship-Level Coding in a 27B Dense Model, Alibaba Cloud Community, 2026.
5. MiniMax M3.1 Flash Weights Are Out: 428B, 252 GB, and a $20M Licence Line, OrcaRouter, 2026.
6. DeepSeek-V4.1-Flash-1017 model card, DeepSeek on Hugging Face, 2026.
7. DeepSeek releases V4.1-Flash, says it outperforms flagship V4-Pro, SiliconANGLE, 2026.
8. Introducing Muse Spark Open, The Most Capable US Open-Weight Model to Date, Meta, 2026.
9. Meta unveils Muse Spark Open, the new top U.S. open weights model, VentureBeat, 2026.
10. Moonshot releases Kimi K3.1 Pro and says it cut its hacking skills, The Next Web, 2026.
11. Cohere says Alibaba's Qwen model tried to send a file to Beijing, TechCrunch, 2026.