Open-source AI models: what is new

admin
Open-source AI models: what is new cover graphic

Open models are AI models whose files you can download and run yourself, or use through a host of your choice. In 2026 the pace of releases has been fast, and the most notable recent one came from Meta in October. At the same time, the word “open” covers very different licence terms, which is where businesses can get caught out.

This article summarises what the sources we read say about open-weight releases from early 2026 to late October 2026, explains the difference between open weights and open source, and gives a short checklist for choosing a model. Many release details come from trackers and news sites rather than vendor pages, so we flag that where it matters.

Key takeaways

  • Meta released Muse Spark Open on 15 October 2026: an 816 billion parameter mixture-of-experts model with 42 billion active parameters per token, under a new Muse Model License.
  • Meta published the benchmark scores itself, and the sources we read say they had not been independently confirmed at launch.
  • A tracker lists dozens of open-weight releases in September and October 2026, from companies including NVIDIA, Google, Mistral, Cloudflare and Xiaomi, but its licence column is often blank.
  • Open weights are not the same as open source: you can often download the model without getting the training data or code.
  • Always read the licence for the exact model version you plan to use, and test it on your own tasks.

Summary

Open-weight releases in 2026 came from many sources. Alibaba’s Qwen line, DeepSeek and others released models through the spring, according to a half-year retrospective. In October, Meta returned to open weights with Muse Spark Open. A release tracker shows a steady run of smaller and specialised models through September and October.

The practical lesson is not which model is biggest. It is that licence terms, hardware needs and independent evidence differ widely, and you should check all three.

Open weights versus open source

An article on model licences, dated 5 September 2026, draws a useful line. Open weight means the trained parameters can be downloaded and run, while training code, training data or documentation may be withheld. Open source, under the Open Source Initiative’s definition as the article describes it, also includes enough information about data, code and parameters for a skilled person to rebuild a substantially equivalent system. The article treats a vendor’s “open source” label as a marketing claim to verify against the licence text.

For a buyer the difference matters in three ways: whether you can inspect how the model was made, whether you can legally use it commercially at your size, and what happens if the vendor changes terms in the next version. Our primer on open-source AI models and who they are for covers the basics.

Earlier in 2026

A retrospective from Digital Applied, published in May 2026, lists these open-weight releases for the first half of the year.

Date Vendor Release Licence per the article
16 February 2026 Alibaba (Qwen) Qwen 3.5 open-weight models Apache 2.0
2 March 2026 Alibaba (Qwen) Qwen 3.5 small models (0.8B, 2B, 4B, 9B) Apache 2.0
16 and 22 April 2026 Alibaba (Qwen) Qwen 3.6-35B-A3B and Qwen 3.6-27B Apache 2.0
24 April 2026 DeepSeek V4 Preview: V4-Pro and V4-Flash Not stated

The article also says Meta released no new open-weight Llama family through 16 May 2026, and that Meta’s closed Muse Spark line launched on 8 April 2026. It notes an inconsistency over the Qwen 3.6-27B date, so verify dates on the model repository before relying on them.

Meta Muse Spark Open

What was released

Two sources agree that Meta released Muse Spark Open weights on 15 October 2026. Unite.AI describes a natively multimodal mixture-of-experts model with 816 billion total parameters, 42 billion active per token and a one million token context window. It accepts text, image, video and audio input and produces text output.

The licence

Both sources describe a new Muse Model License dated 15 October 2026. The terms they report include:

  • Use, copying, modification including fine-tuning, derivative works, distillation and distribution are permitted, subject to conditions.
  • Products must display “Built with Muse Spark Open”.
  • Companies offering AI models as a service, with more than $20 million in gross revenue including affiliates in any consecutive 12 months, must notify Meta and sign a separate agreement. Internal use and use inside your own products are described as exempt.
  • Meta claims no ownership of outputs.
  • A usage policy bars users under 18 and prohibits illegal activity, malware, weapons development, critical infrastructure operation and intentional deception.

Read the licence itself before relying on this summary.

Hardware and availability

Unite.AI reports that running the model locally needs a GPU setup with at least 700GB of VRAM for the NVFP4 weights, or 2TB for BF16, and that Meta says two Apple M5 Ultra Mac Studios can hold the full model. Weights are said to be downloadable from Hugging Face, Amazon SageMaker AI, Google Cloud Model Garden and Microsoft Foundry Models, and hosted by providers including Baseten, Fireworks AI, Groq, Modal, Nebius, OpenRouter and Together AI. For most small businesses, a hosted API is the realistic route.

Performance claims and their limits

Implicator.ai reports that Meta published all benchmark scores in its own model card, and that none had been independently confirmed at launch. In Meta’s table the model trails Kimi K3, GLM 5.3 and Qwen3.8 on a coding benchmark (Terminal-Bench 2.1) and on the Artificial Analysis Intelligence Index, and leads on JobBench and the MCP-Atlas tool use benchmark. Unite.AI says Meta calls it the most capable US open-weight model to date, a claim it attributes to Meta, and that Meta says a fuller evaluation report is coming. Meta’s own stated limitations include possible inaccurate or biased responses and weaker performance in less-supported languages.

How open is it?

Unite.AI notes the model card describes training data only in general terms and does not say the dataset is released. It does not mention released training code. On the definition above, that is open weights, not necessarily open source.

The autumn release wave

The nocode.mba release tracker lists open-weight releases by month. It reports no open-weight releases in August 2026, only closed or platform releases such as Grok 4.6 and GLM-5.3. It lists these for September and October. Because the tracker is a secondary source and often says “not stated” for licences, check each model’s repository.

Date Model Vendor Licence per tracker
20 September Step 5 Preview StepFun Not stated
21 September Qwen-Image-2.1 (image generation and editing) Qwen Not stated
22 September MiMo-V2.6-Pro and Flash Xiaomi Not stated
23 September Nemotron 3 Diarization, NV-Reason-CT and others NVIDIA and partners Mostly not stated
3 October Kolibri Aleph Alpha Apache 2.0
6 October EmbeddingGemma 2 Google Apache 2.0
8 October Mellum2.1 JetBrains Apache 2.0
15 October Muse Spark Open Meta Muse Model License (per other sources)
16 October Beam Reflection Apache 2.0
23 October Kimi K3.1 Pro Moonshot AI Not stated

The tracker also lists Mistral Large 4 as a public preview on 6 October without a licence. It excludes at least one model whose weights it says are only expected later. These are examples, not a complete list. We did not verify these entries against vendor pages.

Licence types to know

The September 2026 licence article groups common terms as follows. It stresses that licence terms change between releases.

Licence type What the article says
Apache 2.0 Permissive and OSI-approved, with a patent licence and no user-count threshold or attribution display requirement
MIT Permissive and OSI-approved, with the main condition being to keep copyright and permission notices
Llama Community License Bespoke and not OSI-approved, with a “Built with Llama” display requirement, naming rules, a separate licence above a monthly user threshold for Llama 4, and carve-outs for some EU use of multimodal Llama 4 models
Other restrictions User or revenue thresholds, territorial carve-outs, acceptable-use limits and commercial-supply bans appear across vendors

The article also notes that hosting a model on a cloud platform does not replace the model’s own licence. This is general information, not legal advice.

What it means for tool buyers

Why consider an open model

Reasons to look at open models include control over where data is processed, the option to fine-tune, and avoiding dependence on one vendor’s pricing. For data handling, running a model inside your own cloud account can simplify privacy answers, though it does not remove your duties. See how to protect company data when using AI tools.

Why be cautious

Big models need serious hardware. Benchmarks published by the model’s maker are claims, not proof. A licence may bar your use case or add duties at your size. And someone has to maintain the deployment, patch it and monitor output quality. If you do not have that capacity, a hosted product may fit better.

Checklist before you build on an open model

  • Read the licence for the exact checkpoint and keep a copy.
  • Check revenue, user and acceptable-use clauses against your plans.
  • Check where it will run and who hosts it, and what data that host sees.
  • Look for independent evaluations, not only the maker’s table.
  • Test it on twenty of your own real tasks. Our guide on comparing AI tools side by side fairly shows how.
  • Plan for a change of licence or model version.

What to watch next

  • Meta’s promised fuller evaluation report for Muse Spark Open, and any independent testing of it.
  • Whether more companies adopt revenue-threshold licences like the Muse Model License.
  • Licence details for models the tracker lists as “not stated”.
  • How open models fare against closed ones in independent evaluations, which we have not reviewed here.

Frequently asked questions

Is open-weight the same as open source?

No. Open weight means you can download the model. Open source, as one source defines it, also means enough information on data and code to rebuild a substantially equivalent system.

Can I use Muse Spark Open commercially?

The sources say many developers can use it free of charge, with a separate agreement needed for companies above a $20 million revenue threshold that offer it as a hosted model service. Read the licence and take advice.

Do I need my own servers to use it?

Not necessarily. Sources list several hosted providers. Running it yourself needs very large amounts of GPU memory according to Unite.AI.

Are open models safer for privacy?

They can give you more control over where data goes, but safety depends on your setup and on the host you choose. See our guide on AI tools and training on your data.

How do I know a benchmark is reliable?

Prefer independent evaluations and your own tests. Scores published by a model’s maker were not independently confirmed in the sources we read.

Conclusion

2026 has been a busy year for open models, with Meta’s October release the headline. Yet the useful questions are plain ones: what does the licence allow, what will it cost to run, and does it do your task well? Start with what open-source AI models are and who they suit, then try a model with a proper trial before committing.

Sources

Keep reading