This open-source LLM could redefine AI research, and it’s 100% public

What is an open-source LLM by EPFL and Eth Zurich

Eth Zurich and the Open-Weight LLM of EPFL offers a transparent alternative to Black-Box IA built on Green Compute and defined for public exit.

The models of large languages (LLM), which are networks of neurons which predict the following word of a sentence, feed the generator of today. Most remain closed, usable by the public, but inaccessible for inspection or improvement. This lack of transparency conflicts with the principles of opening of web3 and innovation without authorization.

Thus, everyone noticed when Eth Zurich and Swiss Federal Institute of Technology in Lausanne (EPFL) announced an entirely public model, formed on the neo-neutral “Alpes” supercomputer in Switzerland and which was planned in Apache 2.0 later this year.

It is generally called “LLM Open in Switzerland”, “a language model designed for the public good” or “the Grand Swiss language model”, but no specific brand or project name has been shared in public declarations so far.

LLM Open – Open Weight is a model whose parameters can be downloaded, audited and locally refined, unlike the “Black -Box” API systems.

Swiss public anatomy LLM

• Ladder: Two configurations, 8 billion and 70 billion parameters, formed on 15 billions of tokens.
• LANGUAGES: Cover in 1,500 languages thanks to a 60/40 non -English English data set.
• Infrastructure: 10,000 NVIDIA grace-hopper chips on “Alps”, entirely fueled by renewable energies.
• License: Open code and weight, allowing the rights to fork and modify researchers and startups.

Which means that LLM Swiss stands out

Switzerland LLM mixes the opening, the multilingual scale and the green infrastructure to offer a radically transparent LLM.

• Architecture open by design: Unlike GPT – 4, which offers only access to the API, this Swiss LLM will provide all its neural network parameters (weight), the training code and the data -based reference references under a APACHE 2.0 license, empowering the developers with fineness, audit and deployment without restrictions.
• Double model sizes: Will be published in 8 billion and 70 billion versions of parameters. The initiative extends slightly to large -scale use with a coherent opening, which GPT – 4, estimated at 1.7 Billion of parameters, does not offer publicly.
• Massive multilingual socket: Trained on 15 billions of token on more than 1,500 languages (~ 60% English, 40% non-English), he questions domination centered on English from GPT-4 with a really global inclusiveness.
• Green, Sovereign Compute: Built on the Alps National Center (CSCS) cluster of Swiss National Center (CSCS), 10,000 NVIDIA grace-hopper superchips offering more than 40 exaflops in FP8 mode, it combines the scale with sustainability absent in private training in the clouds.
• Transparent data practices: In accordance with the protection of Swiss data, copyright standards and EU AI transparency, the model respects OPT counts – without sacrificing performance, highlighting a new ethical standard.

What the AI model has completely opened unlock for web3

The complete transparency of the model allows onchain inference, the data flows to tokenized and the oracle security integrations without required black boxes.

1. ONCHAIN inference: The execution of the guaranteed versions of the Swiss model within Rollup sequencers could allow a summary and fraud tests by intelligent in real time.
2. Tokenized data markets: Since the training corpus is transparent, data contributors can be rewarded with tokens and audited for bias.
3. Composability with DEFI tools: Open weights allow deterministic outputs that oracles can check, reducing the risk of handling when LLMS feed price models or liquidation robots.

These design objectives are properly related to high referencing sentences, including decentralized AI, integration of ia blockchain and onchain inference, increasing the discovery of the article without keywords.

Did you know? Open weight LLM can run inside Rollups, helping intelligent contracts summarize legal documents or indicate suspicious transactions in real time.

IA Market Tailwind that you cannot ignore

• The AI market is expected to exceed $ 500 billion, with more than 80% controlled by closed suppliers.
• Blockchain – AI should go from $ 550 million in 2024 to 4.33 billion dollars by 2034 (22.9% CAGR).
• 68% of companies already manage AI agents and 59% cite the flexibility and governance of the model as criteria for selecting the main ones, a vote of confidence for open weights.

Regulations: the EU law has encountered a sovereign model

Public LLMs, such as the next Swiss model, are designed to comply with the EU AI law, offering a clear advantage in transparency and regulatory alignment.

On July 18, 2025, the European Commission published advice for systemic risk foundation models. The requirements include contradictory tests, detailed training summaries and cybersecurity audits, all from August 2, 2025. Projects of open origin which publish their weights and sets of data can satisfy many of these transparency mandates outside the box, giving public models an advantage of conformity.

Swiss LLM vs GPT – 4

GPT – 4 still has an advantage in raw performance due to the scale and owner refinements. But the Swiss model fills the gap, in particular for multilingual tasks and non -commercial research, while providing auditability that owner models cannot fundamentally.

Did you know? From August 2, 2025, foundation models in the EU must publish data summaries, audit newspapers and opponent test results, the requirements that the next LLM swiss already satisfies.

Alibaba Qwen vs Swiss public LLM: a cross comparison

While Qwen emphasizes the diversity of models and deployment performance, Switzerland LLM focuses on complete transparency and multilingual depth.

Switzerland LLM is not the only serious competitor in the LLM race in open weight. The Qwen series by Alibaba, Qwen3 and Qwen3-Coder, quickly became a very efficient and fully open source alternative.

While Switzerland Public LLM shines with complete transparency, releasing its weights, training code and its entire data methodology, Qwen’s opening is focused on weights and code, with less clarity concerning the training of data sources.

With regard to the diversity of models, Qwen offers an expansive range, including dense models and sophisticated experts in experts in experts (MOE) with 235 billion parameters (22 billion active), as well as hybrid reasoning modes for greater treatment in terms of context. On the other hand, Switzerland Public LLM maintains a more academic objective, offering two clean and research -oriented sizes: 8 billion and 70 billion.

On performance, the Alibaba Coder qwen3 was independently compatible with sources such as Reuters, CIO and Wikipedia Ets to compete with GPT – 4 in coding and stains with high intensity of mathematics. The data on the performance of Switzerland Public LLM is still pending.

On multilingual capacities, Switzerland LLM takes the lead with support for more than 1,500 languages, while Qwen coverage includes 119, still substantial but more selective. Finally, the infrastructure’s imprint reflects divergent philosophies: the LLM public in Switzerland takes place on the CSCS neutral alps supercalcuputor, a sovereign green installation, while Qwen models are formed and served via the Alibaba cloud, priority and the scaling of energy transparency.

You will find below a side by side on the way in which the two LLM Open Source initiatives are measured through the key dimensions:

Did you know? QWEN3-CODER uses a MOE configuration with total parameters of 235b, but only 22 billion are active at the same time, optimizing speed without a complete calculation cost.

Why should manufacturers worry about it

• Complete control: Have the model stack, the weights, code and provenance of data. No supplier or API locking restriction.
• Personalization: Tailor models through domain -specific tasks, onchain analysis, DEFI Oracle validation, code generation
• Cost optimization: Deploy on GPU markets or rolls of rollers; 4 -bit quantification can reduce inference costs from 60% to 80%.
• Compliance by design: Transparent documentation align perfectly with the requirements of the AI EU Act, fewer legal obstacles and deployment time.

Traps to navigate while working with open source LLM

The LLM Open Source offer transparency but face obstacles such as instability, high calculation requests and legal uncertainty.

The main challenges faced by open source LLMS include:

• Scale performance and differences: Despite significant architectures, community consensus wonders if the open source models can correspond to the reasoning, control and integration capacities of closed models like GPT – 4 or Claude4.
• Implementation and instability of components: LLM ecosystems are often faced with software fragmentation, with problems such as version offsets, modules or missing accidents at the time of execution.
• Integration complexity: Users frequently meet dependence conflicts, complex environment configurations or configuration errors when deploying open-source LLMS.
• Intensity of resources: The formation of models, accommodation and inference require substantial calculation and memory (for example, Multi-GPU, 64 GB of RAM), which makes them less accessible to small teams.
• Documentation fallen: The transition from research to deployment is often hampered by incomplete, obsolete or inaccurate documentation, complicating adoption.
• Security and trust risks: Open ecosystems can be sensitive to supply chain threats (for example, typosquat via names of hallucinated packages). Relaxed governance can lead to vulnerabilities such as derivations, inappropriate authorizations or data leakage.
• Legal ambiguities and IP: The use of data refined by the web or mixed licenses can expose users to intellectual property conflicts or violate the terms of use, unlike fully verified closed models.
• Hallucination and reliability problems: Open models can generate plausible but incorrect outputs, especially when refined without rigorous supervision. For example, developers report hallucinated package references in 20% of code extracts.
• Latence and scaling challenges: Local deployments may suffer from slow response time, waiting times or instability under load, problems rarely observed in managed API services.