This roundup is assembled from model cards, vendor announcements and release trackers checked in mid-November 2026. I have not benchmarked these models on my own hardware, so performance claims are attributed and community download figures are quoted from public trackers.
Open-weight models had their biggest month of the year in November 2026, and the headline is simple: the best downloadable model now sits fourth on the main independent leaderboard, behind only closed frontier systems. Kimi K3.1 Pro’s weights went public on 4 November, and nothing else you can download comes close on raw capability.
The catch is that the most capable open-weight models are also the least runnable. A 2.8-trillion-parameter model is open in the legal sense and closed in the practical one for almost everyone reading this. So this guide splits the month into two lists: the frontier releases worth knowing about, and the ones you can actually run on hardware you might own.
Table of Contents
- Where Open-Weight Models Stand in November 2026
- 1. Kimi K3.1 Pro: The New Ceiling
- 2. Mistral Large 4: The Western Open Flagship
- 3. DeepSeek V4.1 Flash: The Value Pick
- 4. GLM-5.3: The Permissive Workhorse
- 5. Qwen3.8 and the Qwen4 Promise
- 6. Gemma 4: The Local Default
- 7. Small Agent Stacks Built on Open-Weight Models
- Open-Weight Models Compared
- Licences Matter More Than Benchmarks
- How to Test Open-Weight Models Before You Commit
- Which Open-Weight Models to Run on Your Hardware
- FAQ About Open-Weight Models
- Verdict
- Sources and Further Reading

Where Open-Weight Models Stand in November 2026
Two facts frame everything else. First, the gap between open-weight models and the closed frontier is small. Epoch AI’s June data put it at roughly four months on its Capabilities Index, a margin that held through the first half of the year. Second, the strongest open models are overwhelmingly Chinese. RedMonk noted in September that, with rare exceptions such as Thinking Machines’ Inkling and NVIDIA’s Nemotron, the largest and best-performing open releases of 2026 came from Chinese labs.
November reinforced both points. Moonshot published Kimi K3.1 Pro in full, Mistral pushed Large 4 out as Europe’s answer, and DeepSeek’s October Flash build became the most downloaded community quantisation of the month.
A quick vocabulary note, because it trips people up. Open-weight models publish their trained parameters so you can download and run them. That is not the same as open source, which would also include training code and data. Almost every model below is open-weight only, and several come with licence conditions that matter for commercial use.
1. Kimi K3.1 Pro: The New Ceiling
Moonshot released Kimi K3.1 Pro through its API on 23 October and shipped the full weights on 4 November, exactly on schedule. It scores 55 on the Artificial Analysis Intelligence Index v5, which puts it fourth among all models and first among downloadable ones.
The architecture is a mixture-of-experts design with about 2.8 trillion total parameters and roughly 104 billion active per token, with a one-million-token context window. That makes it the largest downloadable model available anywhere, according to release trackers.
It is also, for now, a data-centre-only proposition. As of 10 November, no community GGUF conversion had appeared, and even an aggressive quantisation of a model this size needs a multi-GPU server. For most teams the realistic way to use Kimi K3.1 Pro is the API at $2 input and $10 output per million tokens, with the option to self-host later if your volumes justify it.
One design choice stands out. Moonshot says it stripped cyber-related data from both training stages for the open release, keeping offensive security capability below the frontier closed models. For a model that anyone can download, that is a sensible line to draw.
2. Mistral Large 4: The Western Open Flagship
Mistral Large 4, which the company cheerfully calls “Le Chonk”, is a roughly one-trillion-parameter mixture-of-experts model with about 49 to 52 billion active parameters and a one-million-token context. It entered public preview on 6 October and moved to open weights afterwards.
On the Artificial Analysis Intelligence Index v5 it scores 26. That is the highest among Western open-weight models and well below the leading Chinese ones. Its stronger story is coding efficiency: 48% on the Artificial Analysis Coding Agent Index while using 47% fewer tokens per task on average, and Terminal-Bench 4.0 rising from 28.3% in preview to 41% at release.
The licence is the Mistral Ecosystem Open-Weight Licence, a modified MIT licence that is free for internal use but requires a commercial agreement if you offer the model as a hosted service. Mistral’s reference setups assume eight GPUs with tensor parallelism of eight, and there is an NVFP4 checkpoint aimed at Blackwell hardware.
If European data residency or a non-Chinese provenance is a hard requirement, Mistral Large 4 is the obvious choice among open-weight models this month. If it is not, the Chinese releases simply offer more capability per GPU.
3. DeepSeek V4.1 Flash: The Value Pick
DeepSeek’s V4.1 Flash build dated 17 October has been available under MIT weights since that date, and it has quietly become the community’s favourite. A heavily compressed IQ2_XXS GGUF quantisation of the coder build, about 10.1 GB, logged more than 273,000 downloads in nine days after its 2 November upload, according to one release tracker.
That number tells you more than any leaderboard. People are running it, which means it works on real hardware for real tasks. It also appears on SemiAnalysis’s demanding GDPx agent benchmark at 5.4%, ahead of several better-known names, which is respectable for a Flash-tier model.
A warning about that 10 GB file: two-bit quantisations trade a lot of quality for size. They are excellent for trying a model and often fine for code completion, but test carefully before relying on one for anything that needs careful reasoning.
4. GLM-5.3: The Permissive Workhorse
Z.ai’s GLM-5.3 family continues to be the practical choice for teams that want a clean licence. GLM-5.3 Flash ships under MIT, and community builds such as a 13.8 GB IQ2_M GGUF make the larger model approachable on a single high-memory card.
It is not a frontier model. GLM 5.3 sits at 2.7% on GDPx, the bottom of SemiAnalysis’s 13-model table. But that benchmark is brutally hard, and the more relevant question is whether GLM handles everyday agent work at low cost. For many teams it does. We covered the Flash variant in detail in our GLM-5.3-Flash guide, including the one caveat that matters.
5. Qwen3.8 and the Qwen4 Promise
Alibaba’s newest published weights belong to the Qwen3.8 line, with Qwen3.8-Flash-Next the most recent. The community has been busy here too: a 27B Qwen3.8 build at Q8_0, about 30 GB, appeared on 4 November and fits on a single 32 GB card.
The bigger news is Qwen4. Alibaba released Qwen4-Max on 16 November as a closed model, but says smaller Qwen4 models will be open-sourced later. One tracker reports a target of 20 December for Apache 2.0 weights, including a 30-billion-parameter Flash variant with about 3.3 billion active parameters that would need roughly 17 GB at 4-bit. Until those weights are actually downloadable, treat them as a promise, not a product.
6. Gemma 4: The Local Default
Google released Gemma 4 under Apache 2.0, and it remains the safest default for local deployment. The licence is as permissive as they come, tooling support is universal, and the 31B instruction-tuned model fits in about 35 GB at Q8_0, or much less with 4-bit quantisation.
Gemma 4 will not beat Kimi or DeepSeek on hard reasoning. It does not need to. For document summarisation, extraction, classification and private chat on your own machine, it is the open-weight model I would install first, and our best local LLMs guide explains how to size it to your GPU.
7. Small Agent Stacks Built on Open-Weight Models
The most interesting small release of the month was not a model at all. LFM3-Harness, an independent MIT-licensed project released on 1 and 2 November, wires four small LFM3-family models into a local agent stack: a 2B router, a 1.7B actor, a 2B vision model and a 2B voice model. Each is about 1 GB at 4-bit, and the whole thing runs on CPU with llama.cpp servers bound to localhost.
That pattern deserves more attention than it gets. Instead of one giant model, you combine several tiny open-weight models, each doing one job. It runs on a laptop, keeps data on your machine and costs nothing per token. If you want to build one yourself, our guide to a local AI agent walks through the same architecture.
Open-Weight Models Compared
| Model | Size (total / active) | Licence | Runs on | Best for |
|---|---|---|---|---|
| Kimi K3.1 Pro | ~2.8T / ~104B | Check repository licence | Multi-GPU server | Frontier-level reasoning you control |
| Mistral Large 4 | ~1T / 49B to 52B | Modified MIT, commercial licence for hosting | 8-GPU server | European deployments, efficient coding |
| DeepSeek V4.1 Flash | Flash tier | MIT | Single GPU with quantisation | Cheap coding and general use |
| GLM-5.3 / Flash | Varies | MIT (Flash) | Single high-memory GPU | Clean licence, everyday agents |
| Qwen3.8 27B | 27B dense | Check model card | One 32 GB card at Q8 | Strong mid-size local model |
| Gemma 4 31B | 31B | Apache 2.0 | One 24 to 48 GB card | Private local work |
| LFM3 family | ~2B each | Check model cards | CPU or laptop | Tiny local agent stacks |
The table hides one important thing: memory, not compute, is what limits most people. A model’s active parameters decide how fast it runs. Its total parameters decide whether it fits at all.
Licences Matter More Than Benchmarks
If you are building a product, read the licence before the leaderboard. Open-weight models range from genuinely permissive, such as Apache 2.0 and MIT, to custom licences with conditions on hosting, attribution or usage. Mistral’s licence is a good example: free internally, but a separate agreement if you sell the model as a service.
Kimi is a cautionary case. Release trackers disagree about whether its weights ship under Apache 2.0 or a Kimi-specific licence. The model card and licence file in the repository are the only sources that count, so check them before you ship.
The second licence question is export and regulation. Some organisations have policies about models from particular countries, regardless of the licence. That is a business decision, not a technical one, but it is better to know it before your team spends a month fine-tuning. Our comparison of fine-tuning vs RAG is worth reading first, since many teams discover they did not need to fine-tune at all.
How to Test Open-Weight Models Before You Commit
Leaderboards are a starting point. They are a poor finishing point, because the tasks on them are rarely your tasks. Before you standardise on any of the open-weight models above, run a small private test that takes an afternoon.
Pick fifteen to twenty real examples from your own work: support tickets you answered, documents you summarised, code you reviewed. Write down what a good answer looks like for each. Then run every candidate model against the same set, at the same quantisation you plan to use in production, and score the results blind if you can.
Measure three things, not one. Quality is obvious. Speed matters more than people expect, because a model that takes forty seconds per answer will not get used, however clever it is. And memory headroom matters most of all: a model that fits with no room for context will fall over the first time someone pastes in a long document.
Test the quantisation, not the original. Two-bit and three-bit builds behave differently from the full-precision model the leaderboard measured. A 4-bit build of a smaller model often beats a 2-bit build of a bigger one, and you only find out by trying. Our guide to LLM evaluation tools covers frameworks that make this repeatable.
Finally, retest every quarter. Open-weight models improve fast, and the model that lost your test in November may win it by March.
Which Open-Weight Models to Run on Your Hardware
| If you are | Start with | Why | Link |
|---|---|---|---|
| On a laptop with no GPU | LFM3 small models | About 1 GB each at 4-bit, runs on CPU | Hugging Face |
| On a single 24 GB card | Gemma 4 or a quantised DeepSeek V4.1 Flash | Permissive licences and broad tooling support | Ollama library |
| On a 32 to 48 GB card | Qwen3.8 27B at Q8 or GLM-5.3 | Mid-size quality without a server | Best local LLMs |
| With an 8-GPU server | Mistral Large 4 | Official multi-GPU recipes for vLLM and SGLang | Mistral release |
| Needing the best open quality | Kimi K3.1 Pro via API, self-host later | Highest open-weight score, server-class hardware | Kimi on Artificial Analysis |
For serving, pick the engine before the model. Our guide to LLM inference servers covers vLLM, SGLang and llama.cpp, and the right choice depends far more on your hardware than on which open model you pick.
Related on this site: for more context, see DeepSeek V4.1 Flash, Muse Spark Open and Gemma 4 Seven Months On.
FAQ About Open-Weight Models
What are open-weight models?
They are AI models whose trained parameters are published, so you can download, run and fine-tune them yourself. Training data and code are usually not included, which is why “open-weight” is more accurate than “open source”.
What is the best open-weight model in November 2026?
By the Artificial Analysis index, Kimi K3.1 Pro, with a score of 55, fourth among all models. It needs server-class hardware to run locally.
Which open-weight models run on a normal PC?
Gemma 4, quantised DeepSeek V4.1 Flash builds, GLM-5.3 Flash and small LFM3 models all run on consumer hardware. A single 24 GB card covers most of them at 4-bit quantisation.
Are open-weight models free for commercial use?
It depends on the licence. Apache 2.0 and MIT models are broadly free to use commercially. Others, such as Mistral’s, add conditions for hosted services. Always read the licence file.
How far behind closed models are open-weight models?
Epoch AI’s June 2026 estimate put the gap at about four months on its Capabilities Index. November’s Kimi release suggests that gap has not widened since.
When will Qwen4 weights be released?
Alibaba says smaller Qwen4 models will be open-sourced later, and one tracker reports a 20 December target. Qwen4-Max itself is closed.
Verdict
For most readers, the open-weight models worth installing this month are the unglamorous ones: Gemma 4 for private everyday work, a quantised DeepSeek V4.1 Flash for coding, and GLM-5.3 Flash if you want a clean MIT licence. Kimi K3.1 Pro is the most important release, but it is one you will mostly use through an API.
The caveat is that downloadable does not mean deployable. The biggest open-weight models need server hardware, their licences vary more than their marketing suggests, and community two-bit quantisations can lose more quality than download counts imply.
Your next step: install Ollama or llama.cpp, pull Gemma 4 and one quantised DeepSeek build, and run both against ten tasks from your actual week. Keep whichever wins. If neither is good enough, try Kimi K3.1 Pro through its API before buying any hardware.
Sources and Further Reading
- GGUF Discovery’s November dispatch, for release dates, community quantisations and download counts
- Artificial Analysis on Kimi K3.1 Pro, for its index score, speed and pricing
- BrainDetox on Mistral Large 4, for the licence terms and hardware requirements
- RedMonk on open weight models, for the market context and where the strongest releases come from
- AI Central on what open-weight AI means, for the Epoch AI gap estimate and licensing trends
On this site, start with best local LLMs for hardware sizing, then LLM inference servers compared for serving. If you prefer one-click installs, Pinokio covers that route, and local AI agents shows how to put open-weight models to work.


[…] on this site: for more context, see Muse Glimmer, Open-Weight Models in November 2026 and […]