What Are Open-Weight Models? The Latest Releases, and Why They Matter for Sovereign AI

When a new open source LLM lands, the instinct is to check the leaderboard and ask whether it beats last month's model. That question ages fast. Between late July and October 2026, Moonshot AI, Meta, Reflection, NVIDIA, DeepSeek, Z.ai, the Qwen team, Xiaomi and Aleph Alpha all shipped new models, and Mistral has another one queued.

‍

The more useful questions are about licence, hardware and fit. What can you legally do with the model, and can you run it where your data lives? Is it good at your job, and can you swap it out when the next release arrives?

‍

An open source LLM is a large language model whose weights are publicly released for anyone to download, run and adapt; most models sold under that label are, strictly speaking, open-weight models, because their training data and full training code are not published as the Open Source Initiative's definition requires.

What is an open source LLM, and why are most of them actually open-weight models?

In everyday use, an open source LLM is any model whose weights you can download and run yourself. In the strict sense almost none qualify. The Open Source Initiative's Open Source AI Definition (OSAID) 1.0 requires the freedoms to use, study, modify and share a system, and it asks for data information and code as well as parameters.

‍

When the OSI released OSAID 1.0 in October 2024, it checked well-known models against it. OLMo, Pythia, CrystalCoder and T5 met the definition; LLaMA2, Phi-2, Mixtral and Grok fell short. Little has changed: Mozilla's State of Open Source AI report from September 2026 looked at 16 notable open releases and found that none ships the data recipe the definition asks for.

‍

An open-weight model is an AI model whose trained parameters are published for download under a licence, while the data and code used to train it may stay private.

Open source vs open-weight vs closed

• Open source (OSAID 1.0): weights, training code and enough information about the training data to recreate the system, under terms that allow use, study, modification and sharing. OLMo and Pythia are the reference examples.
• Open-weight: downloadable weights under a licence that can range from Apache 2.0 or MIT to a custom licence with attribution rules or revenue thresholds. Training data and recipe usually stay closed.
• Closed: no weights, and often no disclosure of size or architecture. You use the model through the vendor's API.

Why does the label matter if you can download the weights either way? A custom licence can restrict commercial use or impose conditions at scale, and without training-data visibility you cannot fully audit what the model learned.

‍

Regulators also draw lines based on licence terms, as the EU AI Act section below outlines.

Key terms: parameters, active parameters and context windows

Parameters are the numbers a model learns; active parameters are the share of them an MoE model uses for each token; the context window is the most tokens a model can read at once.

• Parameters (weights): the numbers a model learns during training. The parameter count largely decides how much accelerator memory you need.
• Mixture of experts (MoE): a model split into many smaller sub-networks, called experts, plus a router that sends each token to only a few of them. A dense model, such as Qwen3.8-27B, uses all its parameters for every token instead.
• Total vs active parameters: total parameters is the size of the whole MoE model; active parameters is how many are used for each token. Reflection Beam has 501B total but 23B active. Total parameters set the memory you need, because every expert must be loaded; active parameters set the compute per token, and so speed and running cost.
• Token: the unit a model reads and writes, usually a word or part of a word. In English, one token is roughly three-quarters of a word on average.
• Context window: the maximum number of tokens a model can take in at once, covering your prompt, any documents you include and its answer. "1M tokens of context" means about one million tokens, or roughly 750,000 English words. The stated maximum is not always the useful range: Aleph Alpha's Kolibri-1 supports about 1M tokens but recommends 262,144 or fewer.
• Precision and quantisation (BF16, FP8, NVFP4, 2-bit): how many bits store each parameter. Fewer bits mean less memory, usually at some cost in quality. NVIDIA's Nemotron 3.5 Bolt uses a native 2-bit build to fit in 24GB.

What are the recent open-weight model releases?

Since late July 2026, nine labs have shipped new open-weight models, most of them large MoE models with up to 1M tokens of context. The list is grouped by where the developer is based, a procurement question rather than a judgement on quality.

Releases from US labs

• Meta Muse Spark Open (15 October 2026): an 816B total / 42B active MoE with up to 1M tokens of context, taking text, image, video and audio in and producing text. It needs at least 700GB of accelerator memory in NVFP4 and 2TB in BF16, and the Hugging Face repo is gated.
• Meta Muse Glimmer 30B (August 2026): an Apache 2.0 model distilled from Muse Spark for consumer hardware.
• Reflection Beam 1.0 (5 October 2026): a 501B total / 23B active MoE with 1M context, text only, under Apache 2.0. The Hugging Face repo includes BF16, FP8 and NVFP4 builds.
• NVIDIA Nemotron 3.5 Lightning (11 August 2026): a 30B total / 3B active hybrid of Mamba-2, MoE and attention with up to 1M context. It runs on a single H100 80GB at 256K context and ships under the OpenMDW-1.1 licence.
• NVIDIA Nemotron 3.5 Bolt (15 October 2026): a 74B total / 9B active vision-language model that takes text, images and video up to two minutes long, with 256K context, under OpenMDW-1.1.

Releases from Chinese labs

• Moonshot AI Kimi K3 (late July 2026): a 2.8T total / 104B active MoE with about 1M tokens of context, reading text and images, with weights in MXFP4.
• DeepSeek V4.1 Flash (10 September 2026) and its update V4.1-Flash-1017 (17 October): a multimodal MoE on a 552B backbone under MIT, with 8B active parameters per token in prefill and 16B in decode, and up to 1M context (update here).
• Z.ai GLM-5.3: launched 14 August 2026, with weights following about two weeks later, according to NIST's assessment, under the custom GLM-5.3 License.
• Qwen3.8: a 27B dense model under Apache 2.0 (August 2026) with 262K native context, extensible to 1M, plus a 2.4T total / 95B active MoE under the custom Qwen3.8-Max License.
• Xiaomi MiMo-V2.6-Pro (September 2026): a 1.02T total / 42B active omnimodal model under MIT with 1M context.

Releases from European labs

• Aleph Alpha Kolibri-1 (3 October 2026): a 78B total / 3.46B active MoE for German and English under Apache 2.0, with FP8 weights of about 78GB that run on a single H200 or B200, or two H100s.
• Mistral Large 4 (pending): Mistral previewed it on 6 October 2026 as a "1 trillion-parameter natively multimodal model with 52 billion active parameters". As of 27 October, its Hugging Face page lists the weights as upcoming, with an ETA of 31 October and no licence named.

Where Meta's Llama models fit now

Llama is no longer Meta's open model line. Mozilla reports that "Meta retired Llama in April and returned to open weights in August", with the Muse family. The last Meta Llama LLMs, Llama 4 Scout and Maverick from April 2025, remain gated under the llama4 licence.

What do the latest open-weight releases mean?

Read together, the releases show four shifts: MoE with 1M context is the default shape, strong open models come from more countries, open models are a few points behind closed ones, and many licences look open without being open source.

MoE and 1M context are the new baseline

Nearly every large release above is an MoE model, and most are long-context open source LLMs that support up to 1M tokens. The gap between total and active parameters is what makes this practical: Kimi K3 activates 104B of its 2.8T parameters per token, and Kolibri-1 activates 3.46B of 78B.

Strong open models now come from more places

Mozilla's assessment found the strongest open models were "overwhelmingly Chinese", naming Kimi K3, GLM-5.3 and Qwen 3.8. The autumn added strong options from elsewhere. Interconnects calls Muse Spark Open the strongest American-made open model by some margin, and Kolibri-1 gives European buyers an efficient, permissively licensed model from a European developer.

Open models are a few points behind closed ones

On the Artificial Analysis Intelligence Index of 1 September 2026, the best open model trailed the closed leader by three points at 60% of the price, according to Mozilla; a year earlier the leading open model scored 22. For most enterprise workloads, such as retrieval-augmented generation (RAG) over internal documents, a gap that size matters less than fit.

Licences that look open but aren't

Four prominent releases use custom licences. The Kimi K3 License requires a separate agreement for model-as-a-service businesses above US$20M revenue in any 12 months, and "Kimi K3" on the interface of very large products, with internal use exempt. The Muse Model License requires "Built with Muse Spark Open" attribution and a separate agreement for model-as-a-service providers above US$20M gross revenue in any 12 months. The GLM-5.3 License requires a Z.ai security review for model-as-a-service businesses above US$10B revenue.

‍

Calling these models "open source" hides exactly the terms your legal team needs to read.

Which open-weight model is best for coding and other use cases?

GLM-5.3 and Kimi K3 lead for agentic coding, Nemotron 3.5 Bolt suits a single 24GB GPU, and Kolibri-1 covers German and English. Treat these as a shortlist to test: most scores are vendor-reported, and rankings shift monthly.

• Agentic coding: GLM-5.3 and Kimi K3 are the top open-weight models on Datacurve's DeepSWE benchmark, at 69.0% and 68.5% on 113 original software-engineering tasks, with data through September 2026. Kimi K3, at 2.8T total parameters, needs data-centre hardware.
• A single GPU: Nemotron 3.5 Bolt is built to run agentic tasks locally in 24GB, though its benchmarks are vendor-reported. Qwen3.8-27B and Nemotron 3.5 Lightning, on one H100 80GB, are the other compact options.
• Images, video and audio: Muse Spark Open takes all four input types, MiMo-V2.6-Pro is omnimodal, and Bolt handles images and short video.
• German and English: Kolibri-1, a bilingual model from a German developer.
• Least licence friction: Beam, Kolibri-1 and Qwen3.8-27B (Apache 2.0), or DeepSeek V4.1 Flash and MiMo-V2.6-Pro (MIT).

Why do open-weight models matter for enterprises and sovereign AI?

Open-weight models let you run frontier-adjacent AI on infrastructure you choose, at lower cost, without depending on a single vendor's API: the practical foundation of sovereign AI.

Enterprise use is rising

An ETR survey of 200 enterprises, reported by Techstrong.ai, found open-weight models account for 34% of enterprise AI token usage, up from 23% a year ago, and 42% of respondents run them in production. Cost savings are the top reason, and security and compliance reviews are the main blocker; the figures are self-reported, as Constellation Research points out.

Open weights are the practical basis for sovereign AI

Sovereign AI means keeping control over where your AI runs, what data it sees and which vendors you depend on; our sovereign AI explainer covers the concept in depth. When you hold the weights, you can run the model on premises or in a region you pick, keep prompts and documents inside your own infrastructure, and move to another model without renegotiating an API contract.

‍

Choosing sovereign AI models still means looking at where a model comes from. Take a European enterprise seeking AI sovereignty that runs a model from a Chinese lab, such as Kimi K3 or DeepSeek V4.1 Flash, on its own hardware: its data never reaches the model's developer, which is the point of self-hosting.

‍

It still has to test the model's behaviour, read the licence and check whether origin raises procurement or regulatory questions, exactly as it would for a US or European model. Strong options from several regions make that comparison possible.

What the EU AI Act does and doesn't exempt

The EU AI Act gives open models a narrow break, not a blanket one. Article 53(2) exempts general-purpose AI model providers from technical documentation and information for downstream providers when the model is released under a free and open-source licence with public weights, but not for models with systemic risk, presumed above 10^25 FLOPs of training compute.

‍

Copyright and training-data summary duties still apply, and the exemption is lost if the model is monetised, as Arnold & Porter explains. How these rules apply to a specific model is a question for your counsel.

What should you check before running an open-weight model in production?

Check four things before you commit: the licence, the hardware, the model's origin and fit, and how easily you can swap it later. Mozilla and SlashData's survey of 1,411 developers found 79% use open models, but open models reach production 12 points less often than closed ones, and concluded "the gap is tooling, not capability".

• Licence: is it Apache 2.0 or MIT, or a custom licence with attribution, revenue thresholds or vendor reviews that apply to your use?
• Hardware: how much memory do the total parameters need at your chosen precision, and would a smaller model such as Qwen3.8-27B or Nemotron 3.5 Bolt, or a low-active-parameter MoE such as Kolibri-1, cover the task?
• Origin and fit: have you tested the model on your own data and tasks, and does its origin raise policy, procurement or regulatory questions for your organisation?
• Swap-ability: can you replace the model without rebuilding retrieval, prompts and tool definitions?

How do you keep the model swappable with Haystack?

Model swap-ability is mostly an architecture decision. Haystack Enterprise Platform is model-agnostic: the model sits in a pipeline’s LLM or Agent component, so moving from a hosted API to a self-hosted open-weight model means changing that component’s model setting, not rebuilding retrieval or tools. A custom model definition saves your self-hosted model and its parameters for reuse across pipelines, and Prompt Explorer compares up to three pipelines side by side, so you can retune prompts before you switch.

How does harness engineering make an open-weight model perform better?

Harness engineering is the work of designing the code around a model: what it retrieves, how prompts are built, which tools it can call and when it stops. The Meta-Harness paper defines the harness as "the code that determines what information to store, retrieve, and present to the model", and shows one improved harness raising accuracy on hard maths problems by 4.7 points on average across five different models.

‍

This matters most for open-weight models, because the model you can host yourself is often smaller than the frontier model behind an API. A better harness narrows that gap without changing the weights, so compare models inside the harness you will actually run. In Haystack, each part of the harness is a component you can tune and test on its own:

• What the model sees: retrievers and rankers select the few passages that matter instead of filling the context window.
• How it acts: the Agent component calls tools in a loop until an exit condition is met, with a cap on the number of steps.
• How it's checked: a pipeline loop with a ConditionalRouter can send a malformed answer back for another try, and the Agent returns a failed tool call's error to the model so it can recover.

‍

Haystack Enterprise Platform runs the same Haystack components, records a trace of every run so you can see which step failed, is model-agnostic and can be self-hosted in a private cloud or on premises. It includes no model by default: in air-gapped setups you host your own LLMs, embedding and ranking models.

What comes next for open-weight models?

Expect more releases aimed at enterprise and sovereign buyers, of the kind Reflection signalled when it pitched Beam at "enterprises and sovereign nations". Mistral Large 4's weights are the next to watch, and its licence will matter as much as the model. With open models a few points behind closed ones, most of the remaining effort sits outside the model: licence review, security sign-off, serving, harness design and evaluation.

Conclusion

Most open source LLMs are open-weight models. Learn the numbers that describe a release, shortlist models by the job you need done, and judge each one by its licence, hardware needs, origin and how easily you can replace it. Then invest in the harness, because that is where a self-hosted model gains the most.

FAQ

Are Meta's Llama models open source?

No. Llama 4 Scout and Maverick remain gated under Meta's llama4 licence, and the OSI found LLaMA2 fell short of its definition. Meta now releases open weights under the Muse name.

Is DeepSeek open source?

DeepSeek V4.1 Flash is an open-weight model under the permissive MIT licence, so you can download, run and modify it freely. It is not open source in the OSI sense: Mozilla found that none of the 16 notable open releases it reviewed ships the data recipe the definition asks for.

What is the best open source LLM for coding?

GLM-5.3 and Kimi K3 lead open-weight models on Datacurve's DeepSWE coding benchmark. For a single GPU, Nemotron 3.5 Bolt is the compact option to test. Compare your shortlist inside your own agent harness.

What is a long-context open source LLM?

It is an open-weight model that can take in very long inputs, often up to one million tokens at once, roughly 750,000 English words. Kimi K3, Beam, DeepSeek V4.1 Flash and Muse Spark Open all support it, but quality at that length varies, so test with your own documents.