TensorX

What is TensorX?

TensorX is a European AI inference provider serving frontier open-weight models through an OpenAI-compatible API, with inference running in EU data centres and zero data retention by design. 

It gives developers access to model families from DeepSeek, Moonshot AI, MiniMax, Alibaba and Z.ai without requiring teams to operate the underlying inference infrastructure themselves. The service is focused on generative AI and text workloads, with context windows ranging from 131K to 1M tokens and image input available on selected models.

The infrastructure is built around isolated NVIDIA Blackwell GPU clusters located in Dublin, Helsinki and Paris. This provides a European deployment path for teams with data residency and infrastructure isolation requirements. The service is GDPR compliant, including Article 44, and uses ISO 27001-ready infrastructure. 

Its privacy model is based on zero data retention: prompts and outputs are not retained, reused or used for model training. Dedicated isolated GPU clusters are also available for workloads requiring an enterprise SLA.

The catalogue focuses on open weights from leading model labs while separating model choice from inference location. DeepSeek models, Moonshot AI's Kimi family, MiniMax models, Alibaba's Qwen family and Z.ai's GLM models can therefore run on European infrastructure regardless of where their developers operate their own APIs.

This gives teams a practical route to evaluate models such as DeepSeek V4, Kimi K3, MiniMax M3, Qwen3.8 and GLM-5 while keeping inference in Europe. The result is a provider proposition centered on EU residency, long context, open-weight model choice and usage-based inference costs.

TensorX at a glance

TensorX main AI capabilities

• Long-context reasoning: DeepSeek V4 Flash, DeepSeek V4 Pro, MiniMax M3, Kimi K3 and GLM-5.2 each provide a 1M-token context window.
• Agentic and tool-calling workloads: Function calling and tool choice are supported across most models, with specific capability gaps on Qwen3.8 27B, Qwen3 235B A22B and Kimi K2.5.
• Code generation: Kimi K2.7 Code is the coding-focused option, with a 262K-token context window at $1.25/M input and $4.50/M output tokens.
• Vision input: Image input is supported by Kimi K2.5, K2.6, K2.7 Code, K3, MiniMax M3, Qwen3.5 122B A10B, Qwen3.8 27B and GLM-5V Turbo.
• Prompt caching: Prompt caching is supported across the LLM catalogue, with cache reads priced around 75% below fresh input; DeepSeek V4 Flash costs $0.06/M cached versus $0.25/M fresh.

When should you choose TensorX?

Choose TensorX when frontier open-weight capability must stay under EU data residency. Models from DeepSeek, Moonshot AI, MiniMax, Alibaba and Z.ai run on infrastructure in Dublin, Helsinki and Paris. This makes the provider relevant for teams that want access to these model families while keeping inference on European infrastructure with zero data retention and GDPR compliance.

Choose TensorX when long-context processing is central to the workload. Context windows range from 131K to 1M tokens, with DeepSeek V4 Flash, DeepSeek V4 Pro, MiniMax M3, Kimi K3 and GLM-5.2 reaching 1M tokens. These options are relevant for large documents, extended conversations and workflows where substantial context must remain available within a request.

Choose TensorX when frontier API costs are a binding constraint. Pricing spans from $0.07/M input tokens for Qwen3 235B A22B to $3/M for Kimi K3, while output rates vary more sharply. Benchmark lower-cost and mid-tier models against your workload before defaulting to the largest option, because model choice can materially change inference spend.

TensorX pros and cons

TensorX models, features and capabilities on Eden AI

TensorX models are available through Eden AI for Generative AI / Text workloads, with access normalized through the text/chat feature ID.

Relevant selected features for TensorX

TensorX is available on Eden AI under the Generative AI / Text category through the text/chat feature ID. Developers can use the same Eden AI request and response structure to access models from DeepSeek, Moonshot AI, MiniMax, Alibaba and Z.ai. Because Eden AI normalizes the API schema, switching between TensorX models requires changing the model parameter rather than rebuilding the integration. The same approach applies when moving from TensorX to another supported provider, making model comparison, routing and provider changes possible without maintaining a separate API implementation for each inference service.

Popular TensorX models on Eden AI

The following selection highlights well-known TensorX models available through Eden AI across several model families, context sizes and price points.

Explore TensorX models available on Eden AI.

How to choose the right TensorX model

What can you build with TensorX?

TensorX supports applications that combine frontier open-weight models, long context and EU-hosted inference for text, agentic, coding and vision workflows.

Long-context document and codebase analysis

A 1M context LLM API can analyze whole repositories, contract sets or large research corpora without requiring a retrieval step for many workloads. DeepSeek V4 Flash provides a 1M-token context window at $0.25/1M input and $0.30/1M output tokens.

The application can load the relevant source material into the prompt, add analysis instructions and return findings or structured results. When the source fits within the context window, this can remove the chunking, embedding, vector-search and retrieval layers normally required by RAG architectures.

Large prompts still require cost discipline. Teams should evaluate prompt caching and whether every request genuinely needs the complete corpus.

EU-compliant agentic workflows

EU-compliant agentic workflows can combine multi-step reasoning and tool use while model inference remains on European infrastructure. DeepSeek V4 Pro provides a 1M-token context window at $1.75/1M input and $3.50/1M output tokens.

An agent can receive a task, maintain working context, invoke supported tools, process returned data and generate structured results for downstream systems. Inference runs on EU infrastructure in Dublin, Helsinki or Paris with zero data retention by design.

Tool capabilities vary by model, so evaluate the selected model against your agent architecture. Production workflows should also enforce application-level permissions, validation and human approval for consequential actions.

Cost-controlled coding assistants

Cost-controlled coding assistants can match development tasks to models with different token economics. Kimi K2.7 Code costs $1.25/1M input and $4.50/1M output with 262K context, while DeepSeek V3.2 costs $0.30/1M input and $0.50/1M output with 164K context.

A coding application can submit source files, diffs or relevant repository context for code review, refactoring, debugging and test generation. Teams can benchmark multiple models against the same internal tasks to find the appropriate cost-performance point.

Generated code still requires developer review and automated testing, especially for security-sensitive changes, migrations and large refactors.

Multimodal document and screenshot understanding

Multimodal applications can combine image input with text reasoning while inference remains on EU infrastructure. GLM-5V Turbo supports image input with a 203K-token context window at $1.20/1M input and $4.00/1M output tokens.

Applications can submit screenshots, document images or other visual material alongside instructions, then use the generated text for classification, explanation or downstream processing. Image input is also supported by selected Kimi, MiniMax and Qwen models.

Vision performance should be evaluated against representative production data. Important extracted values and visual interpretations should be validated when errors could affect financial, legal, compliance or operational decisions.

TensorX use cases by industry

Best TensorX alternatives and comparisons on Eden AI

The best TensorX alternative depends on whether your priority is model selection, context length, hosting jurisdiction, inference cost or access to closed frontier models.

TensorX vs IONOS

TensorX and IONOS both provide EU-hosted inference for open-weight language models, so European hosting alone does not separate them. The practical differences are model catalogue, infrastructure location and context requirements.

IONOS serves model families including Llama, Mistral, Qwen and gpt-oss from German data centres at standard context lengths. TensorX focuses on DeepSeek, Kimi, MiniMax, Qwen and GLM models, with inference running in Dublin, Helsinki and Paris. Several TensorX options, including DeepSeek V4 Pro, Kimi K3, MiniMax M3 and GLM-5.2, reach 1M-token context windows.

The choice therefore depends on which model families your application needs, how much context it processes and which European hosting location fits your requirements.

Choose IONOS for German-hosted access to Llama, Mistral, Qwen and gpt-oss; choose TensorX for DeepSeek, Kimi, MiniMax, Qwen or GLM with context options reaching 1M tokens.

TensorX vs going direct to the model providers

TensorX and direct model-provider APIs can expose the same open weights, but the inference infrastructure, hosting jurisdiction and commercial terms are different.

TensorX runs supported DeepSeek, Moonshot AI, MiniMax, Alibaba and Z.ai model weights on isolated NVIDIA Blackwell GPU clusters in Dublin, Helsinki and Paris. Its service provides zero data retention by design, GDPR compliance including Article 44, and shared pay-per-token or dedicated cluster options. A direct API from a model developer operates under that developer's own infrastructure locations, data-processing terms, pricing and service conditions.

The model architecture may therefore be the same while the operational environment is not. Teams should compare the specific model version, inference location, retention policy, contract, pricing and service tier rather than treating two endpoints serving the same weights as identical.

Choose the route whose physical inference location and operational terms match your application's data-residency and deployment requirements.

TensorX vs OpenAI and Anthropic

TensorX, OpenAI and Anthropic address different model-selection requirements. TensorX provides open-weight families such as DeepSeek, Kimi, MiniMax, Qwen and GLM on EU infrastructure, with context windows reaching 1M tokens and input prices starting at $0.07/1M tokens.

OpenAI and Anthropic provide their own closed frontier model families. For applications where performance on the hardest reasoning, coding or agentic tasks is the primary requirement, teams should benchmark those frontier models directly rather than assuming an open-weight alternative will match their quality.

Conversely, open weights, EU-hosted inference, long context or lower token costs may make TensorX models better suited to other workloads.

These are complementary requirements rather than interchangeable products. Many teams route different tasks to different model families, and Eden AI provides a normalized API layer for doing that without maintaining separate provider integrations.

Frequently asked questions about TensorX on Eden AI