What is luminous-extended?
luminous-extended was a general-purpose multilingual large language model from Aleph Alpha. The model contained approximately 30 billion parameters, making it the middle-sized model in the original Luminous family. Aleph Alpha positioned it between luminous-base, listed at approximately 13 billion parameters, and luminous-supreme, listed at approximately 70 billion parameters.
The model was built for autoregressive text completion. In practical terms, it generated text by predicting the next token based on the prompt and the preceding context. Its documented task coverage extended beyond open-ended writing: the model was evaluated on classification, closed-book question answering, commonsense reasoning, natural-language inference, and reading comprehension.
luminous-extended was provided by Aleph Alpha, a German AI company whose broader current focus is sovereign, specialized, and enterprise-oriented artificial intelligence. However, this particular model belongs to the company’s earlier Luminous catalog rather than being presented in the supplied sources as a current flagship or actively promoted production model.
Architecture and language support
Aleph Alpha described the Luminous models as decoder-only autoregressive transformers. The architecture used rotary positional embeddings, a technique that helps represent the position of tokens in a sequence. The model was trained on a curated multilingual corpus covering English, German, French, Italian, and Spanish.
Its main distinction was therefore multilingual text generation rather than multimodal interaction. The supplied model record identifies text as both the supported input and output type. Image, audio, video, music, embedding, and other non-text output capabilities are not documented for luminous-extended. It should not be confused with capabilities available elsewhere in Aleph Alpha’s wider platform ecosystem.
What the published benchmarks show
Aleph Alpha’s official benchmark material reported an average score of 51.1 across its core task set and 45.7 across its extended task set for luminous-extended. The evaluation covered classification, closed-book question answering, commonsense reasoning, natural-language inference, and reading comprehension.
The same benchmark material reported that few-shot prompting improved performance. In particular, five-shot prompting brought luminous-extended’s average core-task result close to the zero-shot result reported for luminous-supreme. Few-shot prompting means giving the model a small number of examples in the prompt before asking it to solve a new task.
These figures are provider-published benchmark results from the historical Luminous evaluation, not a current independent ranking. They should be interpreted in the context of the original test design and should not be used as evidence of present-day superiority over newer language models. The benchmark does, however, clarify the intended role of luminous-extended: it was not limited to simple completion, but was also tested on several general language-understanding tasks.
Reasoning, coding, speed, and cost
The model’s documented evaluations included several reasoning-related tasks, including commonsense reasoning, natural-language inference, and closed-book question answering. That supports describing reasoning as an intended evaluation area, but it does not establish a modern reasoning-specialist capability. No chain-of-thought mode, dedicated reasoning setting, or contemporary reasoning benchmark is verified for this exact model.
Coding is not identified as a dedicated capability in the supplied research. The model could potentially generate text that resembles code because it was a general language model, but a coding-optimized training profile, code benchmark, or code-generation feature is not verified. Users selecting a model specifically for software development should therefore look for a currently supported model with documented coding performance instead.
The supplied editorial assessment rates luminous-extended at 4 out of 5 for reasoning, 3 out of 5 for coding, 5 out of 5 for speed, and 4 out of 5 for cost. These are editorial evaluations rather than Aleph Alpha-published specifications. They suggest a favorable speed-versus-capability trade-off within the original Luminous family, but they should not be treated as measured contemporary latency or a verified price comparison.
Availability, pricing, and technical limits
Historical Aleph Alpha material described luminous-extended as available through the Completion Playground and the company’s API client. The supplied research does not verify that the exact model remains accessible today. Its current status is therefore best described as historical or legacy, with ongoing API availability unconfirmed.
No verified contemporary price is available. There is no supported input or output price in the supplied model data, and Aleph Alpha’s current commercial approach is generally oriented toward enterprise and organizational arrangements rather than a clearly advertised consumer subscription for this model.
Several important technical limits are also unverified for luminous-extended:
- The context window or maximum input length is not documented in the supplied sources.
- The maximum output-token limit is not documented.
- A direct knowledge-cutoff date is not available.
- A deprecation date or shutdown date has not been verified.
- Current support for streaming, function calling, structured output, caching, batching, or fine-tuning is unknown.
These gaps matter for production planning. A developer should not assume that a current API accepts the historical model identifier, supports a particular context size, or offers modern response controls simply because the model once appeared in an API client.
Main strengths and limitations
Its clearest strength was the combination of substantial model size and multilingual coverage across five European languages. It was designed for text generation and evaluated on a broad selection of language tasks, making it more suitable for general multilingual experimentation than a narrowly specialized completion system.
The model also occupied a useful middle position in the original Luminous lineup. Compared with a smaller model such as luminous-base, luminous-extended offered a larger parameter count and was intended to provide a higher-capability option. Compared with luminous-supreme, it represented a less resource-intensive middle tier. The supplied research does not provide verified current latency, infrastructure requirements, or prices, so the exact operational trade-off cannot be quantified.
Its main limitations are its uncertain present-day availability and the lack of current technical documentation. There is no verified context length, output limit, price, knowledge cutoff, or modern tool-use specification. It is also not documented as a multimodal model, coding specialist, web-connected system, or function-calling model. These limitations make it difficult to evaluate for a new production deployment.
When to choose luminous-extended
luminous-extended may be relevant when the goal is historical research, reproduction of an earlier Aleph Alpha experiment, or comparison of multilingual European language models. It is also a reasonable subject for benchmarking if an organization already has confirmed access to the original endpoint or a preserved deployment.
Its documented language coverage makes it particularly relevant to experiments involving English, German, French, Italian, and Spanish text completion or classification. Researchers studying few-shot prompting, multilingual question answering, or the evolution of Aleph Alpha’s model family may also find the model useful.
For a new application, choose it only after verifying that the exact model is still available, that the terms of access are acceptable, and that its limits meet the application’s requirements. A currently supported model is likely more appropriate when the project needs a published price, a known context window, reliable service-level expectations, tool or function support, structured output, coding performance, or multimodal input.
Overall assessment
luminous-extended remains a significant historical entry in Aleph Alpha’s original Luminous family: a 30-billion-parameter decoder-only model aimed at multilingual generation and broad language-task evaluation. Its official benchmark record gives useful evidence about its intended capabilities and relative position between luminous-base and luminous-supreme.
It should not, however, be presented as a fully documented current production option. Pricing, context length, output limits, modern API features, and continued access are all unverified. For most new deployments, the model is best treated as a research or archival reference unless Aleph Alpha or an existing deployment confirms current support.