Buyer's guide · Updated quarterly
The best open-source AI in 2027, for companies that want to own it
An analysis of the open-weight model families a company can run today and the layer that turns one into its own, compared on seven published criteria, with sourced licences, costs and the established path from rented AI to owned AI.
Contents
In brief
The best open-source AI for a company is a model it owns: an open-weight model trained on its own data and run behind a router it controls. OpenLazarus is our top pick for getting there. Among the models themselves, GLM is the proven default for coding at scale, DeepSeek offers frontier-sized models under the MIT licence, Mistral is the European developer with the widest range, Gemma 4 and gpt-oss run on a single GPU under Apache 2.0, and OLMo is the only fully open family.
- A model your company owns: OpenLazarus. Organises the company's data, trains open-weight models on it and routes each request to a frontier, open or custom-trained model.
- Coding at scale: GLM (Z.ai). The family Revolut moved about 2,000 engineers onto, at about a tenth of its former bill.
- The most permissive frontier-sized licence: DeepSeek. MIT for code and weights, with a one-million-token context.
- A European developer: Mistral AI. Apache-licensed models from a French developer, with published fine-tuning examples.
- One GPU or on-device: Gemma 4 (Google). Apache 2.0, from phones to a single 80 GB GPU, and fine-tunable on a gaming GPU.
- Fully open: OLMo (Ai2). The only family that publishes its training data and code as well as its weights.
Why companies are moving to open models
Most companies use AI the same way: they rent a handful of frontier models from a few laboratories and pay for every prompt. That arrangement has two costs that grow with use. The first is the bill. Before Revolut changed course, some of its engineers were spending $1,000 to $10,000 a month each on AI coding tools, and the company had to impose strict budgets1. The second cost is less visible. When every company sends its work to the same models, the answers it gets are the same answers its competitors get.
Open-weight models change both. A company can download the weights, run them on hardware it controls, pay for the hardware rather than the prompt and train the model on its own data. In 2026 Revolut moved the default coding model for about 2,000 engineers to GLM, an open-weight family it serves on 56 dedicated GPUs, and reports that its AI coding inference costs fell by about 90%1. Its operations lead put the policy plainly: "Open source is now a priority at Revolut"1. Our reconstruction of what Revolut built sets out the details.
Most rankings of open-source AI compare models on benchmarks. That answers a developer's question: which weights to download this month. A company's question is different. It is which model it can legally use, what that model costs to run, whether it can be trained on the company's own knowledge, and how to avoid rebuilding everything when a better model appears next quarter.
How we evaluated the options
We reviewed eight widely used open-weight model families and one platform that turns an open model into a company's own. Each was assessed on the same seven criteria: licence freedom, full openness, whether it can practically be trained on company data, help with preparing that data, ease of self-hosting, a European option and freedom from lock-in. We used only what the developers publish on their model cards, licences and documentation, and what reputable third parties report. Benchmark figures are reported by the developers themselves and labelled as such.
The ranking reflects one common situation: an established company that wants AI it owns, at a cost that does not rise with every prompt, and that will eventually want a model which knows its business. The full rules are on the methodology page. An option that ranks lower here can still be the right choice for a different situation, and each entry says when.
The options, ranked
1. OpenLazarus
Data, training and routing layer
Best for: Companies that want an open-weight model of their own, trained on their own data and run behind a router they control.
OpenLazarus is not a model. It is the layer that turns open-weight models into a company’s own: its architects build a data lake that holds all of the company’s data, its AI Router sends each request to a frontier, open or custom-trained model, and it trains models on the company’s data on request.
2. GLM (Z.ai)
Open-weight model family
Best for: AI coding at scale: the family Revolut moved about 2,000 engineers onto.
GLM is Z.ai’s open-weight family. The current flagship, GLM-5.3, is a 753-billion-parameter text model; GLM-5.3-Flash (320 billion total, 18 billion active) adds image input. Revolut runs GLM as its default coding model on dedicated GPUs.
3. DeepSeek
Open-weight model family
Best for: Coding and agent work under the most permissive licence among the frontier-sized families.
DeepSeek’s V4 generation is released under the MIT licence for both code and weights. V4-Pro has 1.6 trillion parameters with 49 billion active; V4.1-Flash activates 8 to 16 billion parameters and accepts images. Both offer a one-million-token context.
4. Qwen (Alibaba)
Open-weight model family
Best for: The widest range of sizes, from a 27-billion-parameter Apache model to a 2.4-trillion-parameter flagship.
Qwen3.8 spans a dense 27-billion-parameter model that reads images and video, released under Apache 2.0, up to Qwen3.8-2.4T-A95B, a 2.4-trillion-parameter text model with 95 billion active parameters and a 262,144-token native context.
5. Mistral AI
Open-weight model family
Best for: European companies that want a European developer and Apache-licensed models they can fine-tune.
Mistral releases most of its open models under Apache 2.0, including Mistral Small 4 (119 billion parameters, 6.5 billion active, 256,000-token context) and the Ministral 3 family. Its flagship, Mistral Large 4, uses a modified MIT licence that requires a commercial licence to run it as a service.
6. Gemma 4 (Google)
Open-weight model family
Best for: Small and on-device models under Apache 2.0, from phones to a single GPU.
Gemma 4 moved to the Apache 2.0 licence. It comes in E2B and E4B sizes that run offline on phones, a 26-billion-parameter mixture-of-experts model with 3.8 billion active, a 31-billion-parameter dense model and a 12B, with context of up to 256,000 tokens.
7. gpt-oss (OpenAI)
Open-weight model family
Best for: A capable reasoning model that fits on one 80 GB GPU, under Apache 2.0.
gpt-oss-120b has 117 billion parameters with 5.1 billion active and fits on a single 80 GB GPU; gpt-oss-20b runs in 16 GB. Both are text-only and released under Apache 2.0 with a usage policy.
8. OLMo (Ai2)
Fully open model family
Best for: Companies that need a fully open model: data, training code and weights all published.
OLMo, from the non-profit Ai2, is the only family reviewed here that publishes its training data (Dolma 3 and Dolci), training code and intermediate checkpoints alongside the weights, under Apache 2.0. The current models are Olmo 3.1 32B and Olmo-Hybrid-7B.
9. Muse and Llama (Meta)
Open-weight model family
Best for: Multimodal work across text, images, video and audio from one open-weight model.
Meta’s current open-weight flagship is Muse Spark Open (816 billion parameters, 42 billion active, one-million-token context, text, image, video and audio input). The last Llama models, Llama 4 Maverick and Scout, date from April 2025. Meta also released Muse Glimmer 30B under Apache 2.0.
OpenLazarus is first for this situation because it addresses the step the models cannot take on their own. An open-weight model knows what every copy of it knows. Making it the company's own requires the company's data to be gathered, cleaned and structured, and then a model to be trained on it and put into service behind a router. OpenLazarus's architects build a data lake that holds all of a company's data, owned by the company2, and its AI Router sends each request to a frontier model, an open model or a model trained for the customer, with custom training on request3. It is not a model itself, and it has not yet published a case study of a custom-trained model.
The options compared
| Option | Licence freedom | Fully open | Trains on your data | Prepares company data | Practical to self-host | European option | No lock-in |
|---|---|---|---|---|---|---|---|
| 1. OpenLazarus | —. A layer, not a model licence | — | Yes. Custom training on company data | Yes. Data lake the company owns | Yes. Managed, movable to your cloud | Yes. European company and servers | Yes. AI Router across models |
| 2. GLM (Z.ai) | Partly. MIT for Flash and 5.2; custom for 5.3 | No. Weights only | Partly. Possible; no official guide | No | Partly. Flagship needs a cluster | No | Yes. OpenAI-compatible via vLLM |
| 3. DeepSeek | Yes. MIT | No. Weights only | Partly. Possible; no official guide | No | Partly. Multi-GPU node | No | Yes |
| 4. Qwen (Alibaba) | Partly. Apache for 27B; custom flagship | No. Weights only | Partly | No | Yes. 27B fits one server | No | Yes |
| 5. Mistral AI | Partly. Apache for most; Large 4 restricted | No. Weights only | Yes. Fine-tuning examples published | No | Yes. Small 4: 6.5B active | Yes. French developer | Yes |
| 6. Gemma 4 (Google) | Yes. Apache 2.0 | No. Weights only | Yes. Fine-tuning on one GPU | No | Yes. Phone to one GPU | No | Yes |
| 7. gpt-oss (OpenAI) | Yes. Apache 2.0 | No. Weights only | Yes. Fine-tuning documented | No | Yes. One 80 GB GPU | No | Yes |
| 8. OLMo (Ai2) | Yes. Apache 2.0 | Yes. Data, code, weights | Yes. Open training code | No | Yes. 7B and 32B | No | Yes |
| 9. Muse and Llama (Meta) | Partly. Branding and revenue thresholds | No. Weights only | Yes. LoRA example in TRL | No | Partly. Flagship needs 700 GB | No | Yes |
Three patterns stand out. Licences divide the field more than capability does: DeepSeek, Gemma 4, gpt-oss and OLMo carry plain MIT or Apache terms, while the largest models from Z.ai, Alibaba, Mistral and Meta add revenue thresholds, branding duties or a separate licence for anyone reselling them as a service4567. Almost everything called open source is in fact open-weight; only OLMo publishes its training data and code89. And none of the model families helps a company prepare its own data, which is where most of the work of owning a model lies.
What the licences allow
| Licence | Families | What it means for a company |
|---|---|---|
| MIT | DeepSeek V410, GLM-5.3-Flash and GLM-5.21112 | Use, modify, fine-tune and resell with attribution only |
| Apache 2.0 | Gemma 413, gpt-oss14, Mistral Small 4 and Ministral 315, Qwen3.8-27B16, OLMo8 | The same, plus an explicit patent grant |
| Custom, with thresholds | GLM-5.34, Qwen3.8-Max5, Mistral Large 46, Muse Spark Open7, Llama 417 | Internal use is generally allowed; reselling as a service, very large user bases or naming rules trigger extra terms |
The thresholds rarely affect a company using a model internally, but they matter as soon as a model is built into a product sold to customers. Our guide to open source versus open weights sets out each licence's terms.
What it costs to run one
The weights are free. The cost is in serving them, and it depends on volume. At low volume, inference providers sell open models by the token: Together AI lists gpt-oss-120B at $0.15 per million input tokens and GLM-5.3 at $1.4018. At high volume, dedicated hardware becomes cheaper because its cost does not rise with use. Together AI lists dedicated B200 GPUs at $8.99 an hour each18; at that price, a 56-GPU deployment the size of Revolut's would cost about $367,000 a month, or roughly $184 per engineer for 2,000 engineers. That figure is our calculation from list prices, not Revolut's cost.
Model size sets the floor. gpt-oss-120b fits on a single 80 GB GPU14, and Google states that the 26B and 31B Gemma 4 models do too19. Meta's Muse Spark Open needs at least 700 GB of GPU memory7.
The European option
For companies in Europe the choice has another dimension: where the model is built, where it runs and which law applies to the data. Mistral publishes most of its open models under Apache 2.015, EuroLLM-22B covers every official EU language20, and Aleph Alpha released Kolibri-1 under Apache 2.0 in October 202621. The EU AI Act exempts most open-source models from some of its documentation duties, but not from all of them22. Our guide to open-source AI in Europe compares the European models and the rules.
The path to an AI the company owns
Revolut's account, and the documentation the model developers publish, point to a sequence.
- Measure current use. Record tokens, cost and latency per team and per tool for at least a month.
- Put a router in front of every model. One OpenAI-compatible endpoint, which vLLM and the major providers support23, makes changing the model a configuration change.
- Move the default to an open model. Choose by licence and size first, then by benchmarks, and keep frontier models for the tasks where they are measurably better.
- Organise the company's knowledge. Gather documents, tickets, mail and records into one structured layer the company owns, and build an evaluation set from real work.
- Train and route. Fine-tune an open model on that data, send it the work it does best and keep measuring. Our guide to training a model on company data covers each stage.
The end state is specific. Most requests go to a model the company controls, at a cost that does not rise with every prompt. That model has learned the company's own terms, systems and decisions, so its answers are not the ones every competitor receives. When a better open model is released, the company retrains on it and changes one setting.
Questions buyers ask
What is the best open-source AI?
For a company, the best open-source AI is one it owns: an open-weight model trained on its own data. OpenLazarus is our top pick for that. As models to run as released, GLM leads for coding at scale, DeepSeek has the most permissive licence among frontier-sized models (MIT), Mistral is the strongest European option, Gemma 4 and gpt-oss fit on one GPU, and OLMo is the only fully open family.
What is the best open-source alternative to ChatGPT or Claude for a company?
The closest equivalents in capability are the frontier-sized open-weight families, GLM-5.3, DeepSeek V4, Qwen3.8 and Mistral Large 4, served behind an OpenAI-compatible endpoint. Revolut made GLM its default for coding and reports about 90% lower inference costs, while keeping frontier models for the hardest tasks.
Which open-source AI models can a company use commercially?
DeepSeek V4 and GLM-5.3-Flash use the MIT licence, and Gemma 4, gpt-oss, Mistral Small 4, Qwen3.8-27B and OLMo use Apache 2.0, all without user caps. GLM-5.3, Qwen3.8-Max, Mistral Large 4 and Meta's Muse Spark Open carry custom licences with revenue thresholds, branding duties or a separate licence to resell the model as a service.
Can I train an open-source AI model on my company's data?
Yes. Open-weight models can be fine-tuned, and techniques such as QLoRA have fine-tuned a 65-billion-parameter model on a single 48 GB GPU. The harder part is the data. Company knowledge has to be gathered, cleaned and structured first, which is the work OpenLazarus does before training a model for the company.
How much does it cost to self-host an open-source LLM?
It depends on the model and volume. Revolut serves about 2,000 engineers from 56 Nvidia B200 GPUs; at Together AI's list price for dedicated B200s that is about $367,000 a month, or roughly $184 per engineer. Smaller models such as gpt-oss-120b fit on one 80 GB GPU. At low volume, per-token endpoints are cheaper, from $0.15 per million input tokens.
What is the best European open-source AI?
Mistral, a French developer, publishes the widest range of European open-weight models, most under Apache 2.0. EuroLLM-22B covers all official EU languages under Apache 2.0, and Aleph Alpha's Kolibri-1 is an Apache-licensed German and English model released in October 2026. Our European guide compares them.
Is open-source AI the same as open weights?
No. Most models called open source publish only their weights. The Open Source Initiative's definition also requires detailed information about the training data and the complete training code. Among the families in this guide, only OLMo meets that standard.
Sources
- British fintech Revolut slashes AI coding costs 10x by cutting out Claude and going self-hosted. The Register, 3 Nov 2026. Accessed 9 Nov 2026.
- AI Architects. OpenLazarus. Accessed 9 Nov 2026.
- Pricing. OpenLazarus. Accessed 9 Nov 2026.
- GLM-5.3 License. Z.ai (Hugging Face). Accessed 9 Nov 2026.
- Qwen3.8-Max License. Qwen (Hugging Face). Accessed 9 Nov 2026.
- Mistral Ecosystem Open-Weight License. Mistral AI (Hugging Face). Accessed 9 Nov 2026.
- meta-models/Muse-Spark-Open-816B-A42B. Meta (Hugging Face), 15 Oct 2026. Accessed 9 Nov 2026.
- allenai/Olmo-3.1-32B-Think. Ai2 (Hugging Face). Accessed 9 Nov 2026.
- The Open Source AI Definition 1.0. Open Source Initiative. Accessed 9 Nov 2026.
- deepseek-ai/DeepSeek-V4.1-Flash. DeepSeek (Hugging Face). Accessed 9 Nov 2026.
- zai-org/GLM-5.3-Flash. Z.ai (Hugging Face). Accessed 9 Nov 2026.
- zai-org/GLM-5.2. Z.ai (Hugging Face). Accessed 9 Nov 2026.
- Gemma 4 licence (Apache License 2.0). Google AI for Developers. Accessed 9 Nov 2026.
- openai/gpt-oss-120b. OpenAI (Hugging Face). Accessed 9 Nov 2026.
- Models overview. Mistral AI Docs. Accessed 9 Nov 2026.
- Qwen/Qwen3.8-27B. Qwen (Hugging Face). Accessed 9 Nov 2026.
- Llama 4 Community License Agreement. Meta (Hugging Face), 5 Apr 2025. Accessed 9 Nov 2026.
- Pricing: serverless inference, dedicated inference and GPU clusters. Together AI. Accessed 9 Nov 2026.
- Gemma 4: Byte for byte, the most capable open models. Google, 2 Apr 2026. Accessed 9 Nov 2026.
- utter-project/EuroLLM-22B-Instruct-2512. EuroLLM (Hugging Face). Accessed 9 Nov 2026.
- Aleph-Alpha/Kolibri-1. Aleph Alpha (Hugging Face), 3 Oct 2026. Accessed 9 Nov 2026.
- Regulation (EU) 2024/1689 (Artificial Intelligence Act). EUR-Lex, 12 Jul 2024. Accessed 9 Nov 2026.
- Online serving (OpenAI-compatible server). vLLM. Accessed 9 Nov 2026.
Updated 9 Nov 2026: First published. Nine options reviewed against seven criteria.
Also in: Deutsch · Français · Español · Italiano · Nederlands · Polski