The line-up of LLMs you can run on your own machine changed a lot in 2026. Google shipped Gemma 4, Alibaba shipped Qwen3.8 and DeepSeek shipped V4, while Llama, once the default choice, has not had a new open release at all. This article sorts the main open-weight models as of the end of October 2026 by the question that matters first for local use: will it run on my hardware?
Benchmark numbers are left out on purpose, because almost all of them are self-reported by the vendors. Instead the article sticks to what each model is like, its size, its licence, and roughly how much memory it needs.
Three things to check before choosing
The first filter is not ranking but memory. Quantised to about 4 bits, a model needs roughly "parameters (in billions) × 0.6" GB. A 27B model lands around 16GB, plus a cache that grows with the length of the conversation.
The second is dense versus MoE (mixture of experts). A dense model runs every parameter for every token it generates. An MoE model only uses a few of its many "experts" per token, so its compute cost follows the number of active parameters. "35B-A3B" means 35B parameters in total, of which 3B work on each token. You still need memory for all 35B, but generation runs at close to the speed of a 3B model. If you have memory to spare and want speed, pick MoE; if memory is tight and you want the most capability per gigabyte, pick dense.
The third is the licence. Apache 2.0 and MIT allow almost any use, including commercial; custom licences come with conditions. Sizes within one family sometimes ship under different licences, so check the model card of the exact weights you download.
Models that run on a personal machine
Qwen3.8-27B and Qwen3.6-35B-A3B (Alibaba)
Qwen is the name that comes up most for personal hardware right now. Qwen3.8-27B, released on 14 August 2026, is a 27B dense model that also accepts images. Its native context is 262,144 tokens, extensible to 1,000,000. At 4 bits it fits a 24GB-class GPU or a 32GB Mac.
If speed matters more, Qwen3.6-35B-A3B from April is the alternative. With about 3B active parameters it answers noticeably faster than a dense model of similar quality, as long as you have around 24GB of memory.
Qwen3.8 also has a 2.4T-A95B model, which is nowhere near personal-hardware size. Licences are reported to differ by size, so read the licence file on each Hugging Face repo before using one.
Gemma 4 (Google)
Gemma 4, released in April 2026, is the first Gemma generation distributed under Apache 2.0. It comes in four sizes with clearly different jobs:
- E2B and E4B: small models aimed at phones and boards like the Raspberry Pi. They take images, video and audio, with a 128K context.
- 26B MoE: 26B in total, about 3.8B active. A good balance of speed and capability.
- 31B dense: the most capable of the family. Both larger models support up to 256K context.
Native audio input on the small models is something no other family here offers. For a model that stays resident on a laptop, E4B is the pick; on a desktop, the 26B MoE or the 31B.
gpt-oss-20b and gpt-oss-120b (OpenAI)
OpenAI's reasoning models from August 2025, licensed under Apache 2.0. Most of the weights ship already quantised to MXFP4 (about 4.25 bits), so the download itself is small. The 20b runs in 16GB of memory and the 120b fits on a single 80GB GPU; both have a 128K context.
As of October 2026 OpenAI has not officially announced a successor. More than a year old, gpt-oss-20b is still the standard reasoning model for 16GB machines.
Models strong in Japanese (Swallow, LLM-jp)
For Japanese text, models that Japanese research institutes have further trained on Japanese data are worth adding to the list.
The Swallow project (Institute of Science Tokyo and AIST) released Qwen3 Swallow (8B, 30B-A3B, 32B) and GPT-OSS Swallow (20B, 120B) on 20 February 2026. They continue training Qwen3 and gpt-oss on Japanese data and are licensed under Apache 2.0. Because the sizes match the base models, swapping gpt-oss-20b for GPT-OSS Swallow 20B is straightforward.
The National Institute of Informatics (NII) released LLM-jp-4.1 on 28 September 2026: 8B, 32B-A3B (MoE) and 33B, all "thinking" models that output their reasoning, under Apache 2.0. The training data and development are domestic, which makes the model's provenance easy to explain.
Models that need server-class hardware
The following models have public weights but will not fit on an ordinary PC. They assume an in-house GPU server or a workstation with a very large amount of memory.
DeepSeek V4 was released as a preview on 24 April 2026 under the MIT licence. Even the smaller V4-Flash has 284B parameters (13B active), and its mixed FP4/FP8 weights exceed 150GB. With only 13B active, it may run at a usable speed on a large-memory machine such as a Mac Studio, provided the memory is there. The larger V4-Pro is 1.6T (49B active) and strictly server territory. Both have a 1M-token context.
GLM-5.2 from Z.ai (formerly Zhipu AI) has 753B parameters, a 1M context and an MIT licence. It is popular for coding, and quantised builds for llama.cpp are available.
Mistral released the weights of Mistral Large 4 on 29 October 2026. It has about 1T parameters; the repository name indicates about 52B active. It ships under the "Mistral Ecosystem Open-Weight license", a modified MIT licence that requires a separate commercial licence for offering the model as an API service. Mistral positions it as an open model strong in coding and cyber defence.
Families with no new releases
Meta's Llama, once synonymous with local LLMs, has had no new open model since Llama 4 (Scout and Maverick) in April 2025. Meta's April 2026 flagship, Muse Spark, is closed and its weights are not published. Existing Llama models remain usable, but there is little reason to pick one in 2026.
Microsoft's Phi-5 has not officially shipped as of October 2026 either. For a small model, Gemma 4 E2B or E4B is the obvious substitute.
Candidates by available memory
A rough guide assuming 4-bit quantisation. Long conversations and long documents add cache on top.
- About 8GB: Gemma 4 E4B, Gemma 4 E2B
- 16GB: gpt-oss-20b, GPT-OSS Swallow 20B, Qwen3 Swallow 8B, LLM-jp-4.1-8B
- 24–32GB: Qwen3.8-27B, Qwen3.6-35B-A3B, Gemma 4 26B MoE and 31B, Qwen3 Swallow 32B, LLM-jp-4.1 32B-A3B and 33B
- 64–128GB: gpt-oss-120b, GPT-OSS Swallow 120B
- 192GB or more: DeepSeek V4-Flash
- Server: DeepSeek V4-Pro, GLM-5.2, Mistral Large 4
If you need one starting point: gpt-oss-20b on a 16GB machine, Qwen3.8-27B with 24GB or more. For mostly Japanese text, try the Swallow variant or LLM-jp-4.1 in the same tier side by side.
Software to run them
The weights need a runtime:
- llama.cpp: runs GGUF-quantised models on CPU and GPU. GGUF builds of major models appear shortly after release.
- Ollama: fetches and starts a model with one command, and doubles as an API server.
- LM Studio: a desktop app for browsing and trying models through a GUI.
- MLX: Apple's machine-learning framework for Apple Silicon; runtimes built on it are the fast option on a Mac.
- vLLM and SGLang: inference engines for servers answering many users at once.
For personal experiments, Ollama or LM Studio is the easiest start.
How current this is
This reflects the state of things on 31 October 2026. The field turns over every few months; even within this article, the defaults of six months ago (Qwen3.6, Llama 4) are already giving way to the next generation. Sizes and licences were checked against official announcements and model cards, but read the model card of the exact weights again before you rely on them.
Main official sources: