# Infercom > EU sovereign AI inference platform powered by SambaNova dataflow architecture. Hosted in Germany. GDPR compliant. AI Act ready. Up to 10x faster inference than GPU alternatives. --- ## Homepage ### European Sovereign AI. Breakthrough Performance. Run frontier open-source AI models with world-class performance and full control over where your data lives. Managed infrastructure, dedicated capacity, or on-premises. - EU Data Sovereignty - GDPR compliant & AI Act ready - EU Hosted - No US CLOUD Act exposure - Up to 10x Faster - Up to 5x more energy efficient ### Hosted Your Way Choose how you deploy: our managed infrastructure for instant access, dedicated capacity for guaranteed performance, or on-premises for complete control. **Managed Infrastructure** (Instant Access) Start building in minutes with our fully managed EU-hosted inference platform. OpenAI-compatible APIs, pay-as-you-go pricing, and access to the latest open-source models - no infrastructure management required. **Dedicated Capacity** (Reserved) Guaranteed performance with reserved compute capacity in EU datacenters. Ideal for production workloads that require consistent throughput and predictable latency on large-scale models. **On-Premises** (Complete Control) Deploy SambaNova's inference stack in your own datacenter for complete data sovereignty and operational control. Same performance, your infrastructure. ### Built for European Organizations An inference platform that delivers true EU sovereignty by default, without compromising on performance or developer experience. Certifications and standards: ISO 27001:2022, GDPR, AI Act Ready, Tier III+ **EU Data Sovereignty** All models hosted in EU datacenters by default with complete data residency in German infrastructure. Your data stays in European jurisdiction with no exposure to the US CLOUD Act or PATRIOT Act. ISO 27001 certified. - EU-hosted by default - GDPR & AI Act compliant - No US PATRIOT Act exposure - Full data sovereignty control **World-Class Performance** SambaNova's dataflow architecture delivers up to 10x faster inference than GPU-based alternatives with up to 5x better energy efficiency. The three-tier memory system and native BF16 precision enable running the largest models at full quality without quantization. - SambaNova RDU technology - Superior energy efficiency - Support for largest models - Optimized inference **Developer Experience** OpenAI-compatible APIs mean you can switch from any existing provider with a single line change. Comprehensive documentation, transparent pay-as-you-go pricing, and support for popular frameworks and SDKs. Get started in minutes, scale to production seamlessly. - OpenAI API compatibility - Extensive model library - Pay-as-you-go pricing - Full documentation **Deployment Flexibility** From self-service API access to fully dedicated racks, scale your AI inference to match your workload. Deploy on our infrastructure or bring it into your own datacenter. - Managed infrastructure - Dedicated rack capacity - On-premises deployment - Seamless scaling ### Powerful Open-Source Models Access the latest frontier models optimized for SambaNova's dataflow architecture. All EU-hosted models run with full data sovereignty in German infrastructure. **MiniMax M2.7 Ultraspeed** (Flagship) Latest 229B parameter frontier model with 192K context, built-in self-critique, and native multi-agent orchestration. 400+ tokens/sec on EU infrastructure. [Details](https://infercom.ai/models/minimax-m27) **gpt-oss-120b** (Workhorse) OpenAI's open-weight 120B parameter model. Production workhorse delivering 713 tokens/sec measured p50 with 128K context. Reliable performance without flagship costs. [Details](https://infercom.ai/models/gpt-oss-120b) **Gemma 4 31B** (Vision + Reasoning) Google's most capable dense open model with frontier-class reasoning, native multimodal capabilities, and configurable thinking mode. 31B parameters, 128K context. [Details](https://infercom.ai/models/gemma4) --- ## Sovereign AI ### GDPR Compliant. Truly Sovereign. Blazing Fast. EU Sovereign AI Inference Platform Run open-source AI models on dedicated European infrastructure - with full GDPR compliance, true EU sovereignty, and up to 10x faster inference. As easy to deploy as the non-compliant alternatives. Powered by SambaNova's dataflow architecture. - Hosted in EU - Made in Luxembourg - Up to 10x Faster Inference - GDPR & AI Act Compliant ### Many Providers Are GDPR Compliant. Few Are Truly Sovereign. A US cloud provider's EU subsidiary can legally follow GDPR processing rules. But the US CLOUD Act still gives American authorities the power to demand your data - regardless of where it's stored. True sovereignty means no foreign jurisdiction applies. **GDPR Compliant by Design** EU-hosted models run on dedicated hardware at Equinix Munich, Germany. Your data never leaves the European Union. Full GDPR compliance is built into the infrastructure - not bolted on after the fact. **Zero Data Retention Policy** Your prompts and outputs are never stored, logged, or used for model training. Unlike major AI providers, we don't learn from your data. What you send stays yours. **No CLOUD Act Exposure** This is where compliance and sovereignty diverge. US-parented providers - even with EU data centers - remain subject to the CLOUD Act. Our infrastructure is operated under EU jurisdiction only. **AI Act Ready** Built for EU AI Act requirements from day one. Transparent operations, auditability, and the documentation you need for regulated AI use cases. ### Performance Without Compromise Sovereignty shouldn't mean settling for less. Our infrastructure is powered by SambaNova's dataflow architecture - purpose-built for AI, not repurposed graphics hardware. - Up to 10x Faster Inference - Than GPU-based alternatives on large language models - Up to 5x More Energy Efficient - Lower carbon footprint per token generated - Curated Frontier Models - Best open-source models, handpicked and optimized SambaNova's Reconfigurable Dataflow Units (RDUs) eliminate GPU memory bottlenecks. Full BF16 precision, no quantization needed, millisecond model switching. ### Get Started in Minutes 1. **Sign Up** - Create your account and get your API key in minutes. 2. **Choose Your Model** - MiniMax M2.7, gpt-oss-120b, and Gemma 4 - hosted in the EU on dedicated infrastructure. 3. **Call the API** - OpenAI-compatible REST API. Use curl, the OpenAI SDK, or any HTTP client. ```python from openai import OpenAI client = OpenAI( base_url="https://api.infercom.ai/v1", api_key="your-api-key" ) response = client.chat.completions.create( model="MiniMax-M2.7", messages=[{ "role": "user", "content": "Hello from Europe!" }] ) print(response.choices[0].message.content) ``` OpenAI-compatible API - use the SDK, curl, or any HTTP client. Just point to our endpoint. ### Need Full Control? Deploy On-Premises. For organizations that need complete infrastructure control - regulated industries, government, defense - we offer managed on-premises deployments. SambaNova racks installed in your datacenter, deployed in 90 days. - Single rack footprint - 10 kW average power - Air-cooled, no liquid cooling - 90-day deployment - Air-gap capable ### Part of a Growing Sovereign AI Ecosystem We work with technology partners, system integrators, and AI consultancies across Europe. Whether you need help with integration, custom solutions, or scaling your AI strategy - our ecosystem has you covered. ### A Company With People, Not Just APIs Infercom isn't an anonymous hyperscaler. We're a team of real people who understand your compliance challenges and performance requirements. When you need help, you talk to humans who know your deployment. - Direct access to engineering and support teams - Dedicated account management for enterprise customers - Based in Luxembourg and Munich - your timezone, your language --- ## Technology ### Revolutionary AI Infrastructure Powered by SambaNova Infercom is powered by SambaNova's dataflow architecture - purpose-built for AI inference, delivering unprecedented performance and efficiency. - Up to 10x Faster Inference - Up to 5x Energy Efficient - 24TB Memory per Rack ### Dataflow vs. GPU Architecture Why purpose-built dataflow beats general-purpose GPUs for AI inference. **SambaNova Dataflow Architecture** (Purpose-built for AI) Purpose-built for AI workloads, creating custom processing pipelines for entire computation graphs while minimizing data movement. - Entire model resident in memory - Data flows through operations without intermediate writes - Operator fusion: hundreds of operations in single kernel - Software-defined hardware optimizes for each workload **Traditional GPU** (General-purpose design) General-purpose design requiring kernel-by-kernel execution creates bottlenecks for AI inference workloads. - Kernel-by-kernel execution creates overhead - Excessive data movement between processor and memory - Memory bandwidth bottleneck limits performance - Underutilization of compute resources ### SN40L Reconfigurable Dataflow Unit Built on TSMC's 5nm process with 102 billion transistors, delivering 10.2 PetaFLOPS of compute capacity at BF16 precision. **Chip Specifications:** - 102B Transistors - 1,040 Cores - 10.2 PetaFLOPS - TSMC 5nm Process **Three-Tier Memory Architecture:** - 520MB SRAM - Ultra-fast on-chip cache - 64GB HBM - High-bandwidth memory - 1.5TB DDR - Massive off-package storage **System Configuration:** - 16 Chips/Rack - 24TB Total Memory - 10kW Power - Air Cooled ### World-Record Performance Independent benchmarks by Artificial Analysis. Performance measured in tokens per second per user for real-world inference workloads. **MiniMax M2.7 Ultraspeed** - 400+ tokens/sec (Flagship - EU Hosted) Latest 229B frontier model with 192K context, built-in self-critique, and native multi-agent support. 56% SWE-Pro. EU Sovereign. [Details](https://infercom.ai/models/minimax-m27) **DeepSeek-R1 671B** - 250 tokens/sec (10x vs GPU) The world's largest reasoning model at unprecedented speed. Up to 10x faster than GPU-based providers. 671B params. Available via Global Model Catalog. **Gemma 4 31B** (Vision - EU Hosted) Google's efficient vision-language model with native image understanding. Perfect for document processing and multimodal applications. EU Sovereign. **gpt-oss-120b** - 700+ tokens/sec (Workhorse - EU Hosted) OpenAI's open-weight 120B parameter model. Production workhorse for high-volume sovereign workloads. EU Sovereign. [Details](https://infercom.ai/models/gpt-oss-120b) ### Sustainable AI Infrastructure Up to 5x better energy efficiency than GPU-based inference. **Lower Power Consumption** Average 10kW per rack versus multiple GPU racks consuming 40-50kW+ for equivalent workloads. Reduced chip count translates to dramatic power savings. **Smaller Footprint** Dramatically reduced physical space, simplified cooling, and lower total infrastructure costs. **Air-Cooled Design** No liquid cooling infrastructure required. Standard air cooling simplifies deployment, reduces maintenance complexity, and lowers operational overhead. ### Advanced Model Capabilities **Massive Model Support** Run models up to 671B parameters on a single rack. Support for Composition of Experts (CoE) systems up to 5 trillion parameters with 100+ expert models simultaneously. **Long Context Windows** Handle up to 256,000+ token context windows on single-node deployment. Massive memory capacity enables document analysis, code generation, and reasoning tasks without truncation. **Millisecond Model Switching** Multiple models resident in memory simultaneously with millisecond switching latency - orders of magnitude faster than GPU systems. Perfect for agentic AI and multi-model workflows. ### European Infrastructure Hosted in Equinix Munich 4 - Tier III+ certified, carrier-neutral datacenter. - Munich-Based Hosting - All data and processing remains within German borders under EU jurisdiction. - No US Jurisdiction - Protection from CLOUD Act, PATRIOT Act, and foreign intelligence access. - AI Act Ready - Infrastructure prepared for EU AI Act requirements and compliance. - Tier III+ Certified - 99.982% uptime guarantee with redundant power and cooling systems. --- ## Pricing ### Transparent Pricing for Sovereign AI Inference From developer-friendly pay-as-you-go to dedicated enterprise capacity, all with EU sovereignty by default. - EUR Native Pricing - No Hidden Fees or Minimums ### How Do You Want Your Inference? Start with our inference service and scale to dedicated capacity or on-premises as your needs grow. All options include a fully managed, OpenAI-compatible API. We handle model deployments, infrastructure, and updates - you just call the endpoint. **Inference Service** - Pay per token, priced in EUR - Developer: Pay-as-you-go. Production-ready rate limits. Per-token pricing in EUR. - Enterprise: Custom pricing. SLA, priority support, custom rate limits, and add-ons. **Dedicated Capacity** - Pay per reserved rack Your own reserved SambaNova racks, fully managed by us. Hosted in our EU datacenters with the same operational simplicity as the inference service, but with guaranteed capacity and priority support. - Guaranteed capacity - Reserved racks, no software rate limits - Custom model hosting - Any SambaNova-supported model - Performance SLAs - Contractual uptime guarantees - Priority support - Direct engineering access **On-Premises** - Own the hardware and software stack Deploy the complete inference stack in your own datacenter. Air-cooled, no liquid cooling required. Start with a single rack and scale as needed, with dedicated 24/7 support. - Complete inference stack - Hardware + software, ready to run - Air-cooled, single rack - No liquid cooling, start with one rack - Unlimited models - Deploy anything SambaNova supports - 90-day deployment - Faster than GPU alternatives ### Compare Features Across Tiers | Feature | Inference Service | Dedicated | On-Premises | |---|---|---|---| | EU-Hosted by Default | Yes | Yes | Your Location | | Pricing Model | Pay-per-token | Reserved capacity | Custom licensing | | Rate Limits | Per plan | Hardware only | Unlimited | | Model Catalog | Standard models | Standard + Custom | Any model | | Custom Model Hosting | - | Yes | Yes | | Support | Docs & Community to Priority | Priority | Dedicated 24/7 | | SLA Guarantee | Best effort to Custom | Custom | Custom | | Air-Gapped Deployment | - | - | Yes | | Data Residency Control | EU default | EU guaranteed | Your choice | | Best For | Prototyping to production | High-volume production | Full physical ownership | ### The Infercom Advantage **Sovereignty Included** EU data sovereignty isn't an add-on or premium feature. It's included by default in every tier, at no extra cost. **Transparent Token Pricing** Clear per-token pricing in EUR with no hidden fees. What you see is what you pay. No surprise charges for data transfer or API calls. **No Performance Throttling** Every request runs at full inference speed regardless of your plan - we never reduce token throughput or deprioritize pay-as-you-go users. Rate limits cap request frequency, not performance. **Clear Upgrade Path** Start with pay-as-you-go, scale to dedicated capacity, deploy on-premises. Move between tiers as your needs evolve. ### Frequently Asked Questions **What are the rate limits?** Rate limits vary by plan and model. The Developer tier has standard rate limits documented at docs.infercom.ai. Enterprise plans offer custom rate limits tailored to your workload. Rate limits control request frequency - they do not affect inference speed per request. **What payment methods do you accept?** For our inference service, we accept major credit cards through Stripe. Enterprise, dedicated capacity, and on-premises customers can also pay by invoice. **Can I use models not in your catalog?** Yes! With dedicated capacity or on-premises, you can run any model supported by the SambaNova platform, including fine-tuned variants. On the inference service, you can request specific models and we'll evaluate adding them to our catalog. **What's included in EU sovereignty?** All data processing happens in our Munich datacenter (Equinix MU4) by default. Your data never leaves EU jurisdiction unless you explicitly opt to use non-EU models. Full GDPR compliance and AI Act readiness included. **Do you offer volume discounts?** Yes - both Enterprise and dedicated capacity contracts offer custom pricing based on your committed usage. Contact our sales team to discuss your requirements. **How quickly can I deploy on-premises?** Our standard deployment timeline is 90 days from contract signing to production-ready, including hardware delivery, installation, and configuration. This is significantly faster than traditional GPU-based alternatives. **What's the difference between inference service and dedicated?** The inference service is our multi-tenant platform where you share resources with other users - available in Developer and Enterprise plans. Dedicated capacity reserves entire SambaNova racks exclusively for your use with guaranteed performance and no software rate limits - only the physical hardware capacity applies. **Can I switch between tiers?** Yes! You can upgrade from the inference service to dedicated capacity at any time. Moving to on-premises requires planning for hardware procurement and installation. Contact us to discuss your migration path. --- ## Performance ### Don't Take Our Word for It Real benchmark results from our EU infrastructure. Independently verified technology. Open-source tools you can run yourself. ### Tokens per Second, Measured These numbers are from our production API with our servers in the EU. Not vendor marketing - real measurements you can reproduce. Last measured: July 2026 Config: 10K input / 1K output, single request, server-side p50 **gpt-oss-120b** (EU Sovereign) - fastest throughput - Output Throughput: 713 tok/s - Time to First Token: 388 ms - End-to-End Latency: 1.789 s - Peak: up to 772 tok/s on shorter prompts - Context window: 128K (131,072 tokens); licence: Apache 2.0 **MiniMax M2.7 Ultraspeed** (EU Sovereign) - 229B frontier reasoning - Output Throughput: 428 tok/s - Time to First Token: 690 ms - End-to-End Latency: 3.023 s - Peak: up to 444 tok/s on shorter prompts - Context window: 192K (196,608 tokens); open-weight **Gemma 4 31B** (EU Sovereign) - native multimodal - Output Throughput: 199 tok/s - Time to First Token: 1189 ms - End-to-End Latency: 6.206 s - Native image + text understanding in one model - Context window: 128K (131,072 tokens); licence: Apache 2.0 Server-side metrics (p50). Measured using our open-source benchmark tool at https://benchmark.infercom.ai. Your client-side results will vary based on network location and conditions. ### Fastest LLM Inference, by Workload Which number decides your speed depends on what you are building. - **Chat and voice -> Time to First Token.** Lowest measured TTFT: gpt-oss-120b at 388 ms (p50). - **Agents and coding -> Output throughput.** Highest measured throughput: gpt-oss-120b at 713 tok/s (p50). For frontier reasoning across long runs, MiniMax M2.7 Ultraspeed measures 428 tok/s with a 192K context window. - **Batch and RAG -> End-to-end latency.** Fastest measured end-to-end: gpt-oss-120b at 1.789 s for a 10K-in / 1K-out request (p50). Gemma 4 31B takes image and text in the same request. ### LLM Inference Speed: Frequently Asked Questions **Who has the fastest LLM inference in Europe?** Infercom serves open-weight models such as gpt-oss-120b on SambaNova dataflow hardware in Munich. Our highest measured output is 713 tokens per second, with peaks up to 772 tok/s on shorter prompts - up to 10x faster than GPU-based alternatives. Every figure is a server-side p50 from our production API, and reproducible with our open-source benchmark tool. **Which model is fastest, and for what?** gpt-oss-120b leads output throughput at 713 tok/s and also has our lowest time to first token at 388 ms. MiniMax M2.7 Ultraspeed is our 229B frontier reasoning model at 428 tok/s with a 192K context window. Gemma 4 31B measures 199 tok/s and adds native image and text input. Chat and voice depend on time to first token; agents and coding depend on throughput. **How is this measured - can I reproduce it?** Yes. Every number is a server-side p50 at 10K input and 1K output tokens, single request, measured against our production API with our open-source benchmark tool. Run it online with your own key, or clone the repository and run it locally. **How do you keep latency low under concurrency?** Dataflow hardware keeps its compute pipeline busy at low batch sizes, so per-request speed does not depend on stacking many requests together the way GPU serving does. The figures published here are single-request p50s; the benchmark tool has a real-workload mode that replays variable request rates. **Is the speed independently verified?** The dataflow architecture Infercom runs on is continuously benchmarked by Artificial Analysis and has been covered by VentureBeat and TechRadar, with the underlying energy-efficiency methodology validated by Stanford Hazy Research. Those sources cover the technology; the per-model numbers here are our own measurements on our own Munich infrastructure. **Why is dataflow faster than GPUs?** A GPU repeatedly moves model weights and activations between compute and memory, and during token-by-token decoding it spends much of its time waiting on that memory traffic. A dataflow architecture maps the model onto the chip and streams data through it, so far more of the silicon does useful work per token. That is where the up to 10x speed advantage over GPU-based inference comes from, alongside up to 5x better energy efficiency. ### What We Measure **Time to First Token (TTFT)** How quickly the model starts responding after your request. Critical for interactive applications and chat interfaces. **Output Throughput** Tokens generated per second after the first token. Determines how fast a complete response is delivered to the user. **End-to-End Latency** Total time from request to complete response. Includes TTFT plus full generation time. The number that matters for batch workloads. ### Run Your Own Benchmark Our benchmark tool is fully open source. Run it against our API with your API key, or clone the repository and run it locally. Same code, same methodology, your results. Benchmark modes: - **Synthetic Performance** - Fixed input/output token counts for controlled comparisons across models - **Real Workload Simulation** - Variable request rates mimicking production traffic patterns - **Custom Dataset** - Upload your own prompts and measure performance on your actual workload - **Interactive Chat** - Per-response metrics in a live chat interface - see TTFT and throughput on every reply ### Performance Without the Power Bill Up to 5x more energy efficient than GPU-based inference. Speed doesn't have to come at the planet's expense. - **10 kW Per Rack** - vs. 40-50 kW+ for equivalent GPU infrastructure - **Air Cooled, No Liquid Cooling** - Standard air cooling simplifies deployment and eliminates water usage - **Up to 5x More Efficient** - More intelligence per joule of energy consumed --- ## Getting Started ### Build with EU Sovereign AI OpenAI-compatible APIs. World-record performance. Complete data sovereignty. Start building in minutes. ### Start in Three Steps 1. **Sign Up & Get API Key** - Create your account and generate your API key instantly. 2. **Choose Your Model** - Access latest open-source models: MiniMax M2.7 Ultraspeed, gpt-oss-120b, Gemma 4, and more. Full 16-bit precision, no quantization. 3. **Make Your First Call** - OpenAI-compatible API means you can use existing tools, libraries, and code. Just change the base URL. ### OpenAI-Compatible API Drop-in replacement for OpenAI. Works with your existing code, tools, and frameworks. - Same API format as OpenAI - Works with LangChain, LlamaIndex, CrewAI - Python, JavaScript, TypeScript, REST ```python from openai import OpenAI client = OpenAI( base_url="https://api.infercom.ai/v1", api_key="your-api-key" ) response = client.chat.completions.create( model="MiniMax-M2.7", messages=[{ "role": "user", "content": "Explain quantum computing" }], temperature=0.7 ) print(response.choices[0].message.content) ``` ### Everything You Need to Build - **OpenAI-Compatible** - Drop-in replacement. Works with existing OpenAI client libraries and frameworks. - **World-Record Speed** - Up to 10x faster inference powered by SambaNova's dataflow architecture. - **Latest Models** - MiniMax M2.7, gpt-oss-120b, Gemma 4, and more. Full precision, regularly updated. - **EU Sovereignty** - Hosted in EU. No data retention. GDPR compliant by design. ### Powered by SambaNova Technology **High-Performance Inference** Built on SambaNova's SN40L RDU with full 16-bit precision - no quantization required. - Up to 10x faster than GPU inference - Full BF16 precision - Optimized for models 70B+ parameters **Model Bundling for Agentic AI** Run multiple models with millisecond switching. Perfect for complex agentic workflows. - Millisecond model switching - 100+ models on single rack - Ideal for multi-agent systems **Developer-Friendly Ecosystem** OpenAI compatibility, AI Starter Kits, and integrations with popular frameworks. ### Built for European Compliance **Hosted in EU** All inference processing happens in European Union datacenters. Your data never leaves EU jurisdiction. **No Data Retention** True transient processing. Prompts and responses are processed but never stored, logged, or used for training. **Regulatory Ready** Built for regulated industries requiring strict compliance with European data protection laws. GDPR & AI Act aligned. ISO 27001 certified. DPA available. --- ## Agentic Coding ### MiniMax M2.7 for Agentic Coding Run MiniMax M2.7 Ultraspeed with your favorite AI coding tools - frontier-level quality at a fraction of the cost, hosted on EU sovereign infrastructure. ### Why Developers Choose MiniMax M2.7 - **400+ tok/s** - Fast enough for real-time agentic workflows with instant tool calls - **13x Cheaper** - Approximately 13x cheaper than Opus at comparable quality - **EU Sovereign** - Your code stays in European jurisdiction. GDPR compliant. ISO 27001 certified. - **Scale Without Surprises** - Predictable EUR pricing, no surprise bills, no rate limit anxiety ### SWE-Pro Performance MiniMax M2.7 Ultraspeed achieves 56.22% on SWE-Pro, matching GPT-5.3-Codex. Built-in self-critique and native multi-agent orchestration make it ideal for complex agentic workflows. ### Two Ways to Use Infercom for Agentic Coding **Option 1: Full Replacement** Use MiniMax M2.7 Ultraspeed for everything - simplest setup, one model, one provider. 56% SWE-Pro matches frontier performance at a fraction of the cost. **Option 2: Planner/Executor Split** Keep your frontier model (Claude, GPT, Gemini) for planning (5-15 turns), use MiniMax M2.7 Ultraspeed on Infercom for execution (50-200+ turns). Best of both - frontier reasoning + fast execution. Tools with native planner/executor support: Codex CLI, Cline, OpenCode, Cursor. ### Why Speed Matters for Coding Agents Coding agents spend 80-95% of their turns on execution - file reads, edits, test runs, retries. A 4x speedup on execution means 3-4x faster overall task completion. A single agentic task can consume 50,000-200,000 tokens across hundreds of iterations. At frontier pricing, that's €3-15 per task. At Infercom rates, it's €0.15-0.60. ### Supported Tools MiniMax M2.7 works with popular agentic coding tools via OpenAI-compatible API: - Cursor - AI-first code editor - Cline - VS Code extension for AI-assisted coding - Codex CLI - OpenAI's terminal coding agent - Aider - Terminal-based AI pair programming - OpenCode - Open-source Claude Code alternative - Continue - Open-source AI code assistant - Windsurf - AI-powered IDE - Kilo Code - VS Code AI coding extension - Claude Code - Anthropic's CLI coding assistant - Goose - Autonomous AI coding agent - Blackbox - AI coding assistant Integration guides: [https://docs.infercom.ai/en/agentic-coding](https://docs.infercom.ai/en/agentic-coding) ### Developer Testimonials "M2.5 gave me the best result I've gotten so far. Better than Claude Code with Opus 4.6. A typical SWE-Bench task costs about $0.15 with M2.5 versus $3.00 with Opus." - Thomas Wiegold "I went from Claude Max ($100/month) to MiniMax ($20/month) with the same usage patterns and haven't hit limits once." - NotChrisGroves ### Trust & Compliance - ISO 27001:2022 Certified - GDPR Compliant - Powered by SambaNova (Germany datacenter) - No code retention - your prompts and outputs are never stored Page: [https://infercom.ai/developers/agentic-coding](https://infercom.ai/developers/agentic-coding) --- ## About ### Europe's First Sovereign AI Inference Provider Building the foundation for European AI sovereignty with world-record performance, complete data control, and unwavering commitment to European values. ### Our Mission To provide European organizations with AI infrastructure that delivers world-class performance without compromising on data sovereignty, regulatory compliance, or European values. ### Our Vision To establish Europe as a leader in sovereign AI infrastructure, demonstrating that organizations don't have to choose between cutting-edge technology and data sovereignty. ### What Makes Infercom Different - **European-First by Design** - Not a European region of a US cloud - built from the ground up for European needs. - **Performance Leadership** - Up to 10x faster inference than GPU alternatives with 5x better energy efficiency, powered by SambaNova's dataflow architecture. - **Absolute Data Sovereignty** - Your data never leaves EU jurisdiction. No US CLOUD Act exposure. - **Compliance-Ready** - Purpose-built for GDPR and AI Act compliance. ISO 27001 certified. - **Partnership Approach** - Hands-on support and technical guidance rather than just infrastructure rental. - **Open & Transparent** - OpenAI-compatible APIs, major open-source models, and no vendor lock-in. ### European Infrastructure | Detail | Value | |---|---| | Datacenter | Equinix Munich (Tier III+) | | Jurisdiction | German / EU (No US CLOUD Act) | | Uptime SLA | 99.982% Guaranteed | | Certification | ISO/IEC 27001:2022 | Certificate No. LU-IS-20250253. Valid 16.12.2025 - 15.12.2028. Certified by Proks Certification GmbH, accredited by DAkkS. ### Our Values - **Sovereignty First** - Data control as a fundamental right, never compromised. - **Performance Without Compromise** - World-leading performance proving Europe can lead in both sovereignty and innovation. - **Transparency & Trust** - Complete transparency about our infrastructure and partnerships. No hidden dependencies. - **European Values** - Privacy, sustainability, democratic governance, and rule of law at the core of everything we do. ### Leadership **Cem Tufekci** - Founder & Chairman Tech entrepreneur and investor. Founded Qualist Technology (2009). Active angel investor across enterprise software and deep tech. **Altug Eker** - Managing Director Deep expertise in AI and cloud infrastructure. Previously at Vodafone, Bertelsmann, and ICT Cloud Computing Services. **Serik Kaldykulov** - Co-Founder Founder & Managing Partner at Elefund. Early-stage investor with a portfolio spanning AI infrastructure and fintech. **Gokay Turanlioglu** - Board Member Background in private equity and M&A. Previously at Merrill Lynch and Esas Holding. ### Investors - Qualist Investment AG - Swiss Venture Capital - Pakua Capital AG - Swiss Single Family Office - Serik Kaldykulov - Solo Investor --- ## Ecosystem ### Building the EU Sovereign AI Ecosystem - Together Partners, technology, and infrastructure united under European governance. From datacenter to application - sovereign at every layer. ### Ecosystem Partners Partners: SambaNova, Cortecs, NeurX, 6away, Midio, Melious, Co-mind, DeepMask, Qualist, OpenGPU, EUrouter, Opper, BasePeak ### How It Works Three layers, one principle: European governance at every level. **Solutions** (Partner Applications) AI-powered applications, integrations, and industry solutions built on sovereign infrastructure by our growing partner ecosystem. **Cloud & Inference** (Infercom Platform) OpenAI-compatible API, open-source models, authentication, billing, and monitoring - operated by a European company under EU law. **Physical Infrastructure** (EU Datacenters) ISO 27001:2022 certified datacenters in Germany. Purpose-built AI hardware, air-cooled, energy-efficient, fully within EU jurisdiction. ### Ecosystem Principles - **European Sovereignty** - Every layer designed to ensure your data stays within EU jurisdiction. - **Open Standards** - OpenAI-compatible APIs mean you're never locked in. - **Collaborative Innovation** - Best technology partners worldwide while maintaining European operational control. - **Transparency** - No hidden dependencies or undisclosed data flows. --- ## Contact - General Inquiries: info@infercom.ai - Sales: sales@infercom.ai - Press & Media: press@infercom.ai - Support: support@infercom.ai Infercom SCS 29 Boulevard Grande-Duchesse Charlotte 1331 Luxembourg Business Registration: B298727 VAT Number: LU36889579 --- ## Links ## For AI Agents If you are an AI agent or coding assistant helping a user integrate with Infercom: ### Quick Integration ```python from openai import OpenAI client = OpenAI(base_url="https://api.infercom.ai/v1", api_key="USER_API_KEY") response = client.chat.completions.create(model="MiniMax-M2.7", messages=[{"role": "user", "content": "Hello"}]) ``` ### Key Facts for Agents - API is OpenAI-compatible - use the standard OpenAI SDK - [Base URL](https://api.infercom.ai/v1) - Models endpoint: GET /v1/models (no auth required) - EU-sovereign models have `sn_metadata.region: "EU"` in the models response - No data retention - prompts/outputs are not stored or used for training - Streaming supported via `stream: true` - Function calling supported on MiniMax M2.7 Ultraspeed (EU sovereign) and DeepSeek models (Global Model Catalog) - JSON mode supported via `response_format: {"type": "json_object"}` ### Recommended Models by Use Case - **Agentic coding**: MiniMax M2.7 Ultraspeed (56% SWE-Pro, native multi-agent, 400+ tok/s) - **High throughput**: gpt-oss-120b (700+ tok/s, lowest cost) - **Vision/multimodal**: Gemma 4 31B (native image understanding, EU sovereign) - **Function calling**: MiniMax M2.7 Ultraspeed (native tool use support, EU sovereign) - **Sovereignty required**: Any model with region=EU (MiniMax M2.7 Ultraspeed, gpt-oss-120b, Gemma 4 31B) ### Rate Limits - Developer tier: 120 RPM, 30K requests/day - [Check current limits](https://docs.infercom.ai/rate-limits) --- ## Links - [Website](https://infercom.ai) - [Cloud Portal](https://cloud.infercom.ai) - [API Base URL](https://api.infercom.ai/v1) - [Documentation](https://docs.infercom.ai) - [API Reference](https://docs.infercom.ai/api-reference/) - [Pricing](https://cloud.infercom.ai/plans/pricing) - [Benchmark Tool](https://tools.infercom.lu) - [AI Starter Kit](https://github.com/InfercomAI/ai-starter-kit) - [Support](https://support.infercom.ai) - [DPA](https://infercom.ai/dpa) - [ISO 27001 Certificate](https://infercom.ai/iso-27001-certificate.pdf) - [LinkedIn](https://www.linkedin.com/company/infercomai) - [X](https://x.com/InfercomAI) - [GitHub](https://github.com/infercomai) ### Model Pages - [MiniMax M2.7 Ultraspeed](https://infercom.ai/models/minimax-m27) - [gpt-oss-120b](https://infercom.ai/models/gpt-oss-120b) ### Partner Pages - [BasePeak](https://infercom.ai/partners/basepeak) - [Co-Mind](https://infercom.ai/partners/co-mind) - [CoformAI](https://infercom.ai/partners/coformai) - [DeepMask](https://infercom.ai/partners/deepmask) - [Melious](https://infercom.ai/partners/melious) - [Neur-X](https://infercom.ai/partners/neur-x) ### Developer Resources - [Agentic Coding](https://infercom.ai/developers/agentic-coding) - [Agentic Coding Docs](https://docs.infercom.ai/en/agentic-coding) ### Glossary - [Inference Performance Glossary](https://infercom.ai/glossary) - [Open-Weight Model](https://infercom.ai/glossary/open-weights-model) - [Context Window](https://infercom.ai/glossary/context-window) - [Parameters](https://infercom.ai/glossary/parameters) - [Data Residency](https://infercom.ai/glossary/data-residency) - [Data Processing Agreement (DPA)](https://infercom.ai/glossary/dpa) - [GDPR for AI Inference](https://infercom.ai/glossary/gdpr) - [Zero Data Retention (ZDR)](https://infercom.ai/glossary/zero-data-retention) ### Blog (Insights) - [Blog Index](https://infercom.ai/blog) - [Open-Weight AI Models Explained](https://infercom.ai/blog/open-weight-models-explained) - [How Dataflow Delivers Speed](https://infercom.ai/blog/how-dataflow-delivers-speed) - [LLM Inference Speed Explained](https://infercom.ai/blog/llm-inference-speed-explained) - [Token Pricing Fallacy](https://infercom.ai/blog/token-pricing-fallacy) - [Inference Speed in Agentic Coding](https://infercom.ai/blog/agentic-coding-speed) --- ## Blog: How Dataflow Delivers Speed ### 713 Tokens Per Second: The Architecture Behind Ultraspeed Infercom delivers 713 tokens per second on gpt-oss-120b and 428 tok/s on MiniMax M2.7 Ultraspeed from our EU infrastructure. You can reproduce these numbers with our open-source benchmark tool at benchmark.infercom.ai. But speed claims are meaningless without understanding why. We're hardware agnostic. Our job is to evaluate inference architectures and deploy the best option for each workload. For memory-bound decode on large models, we currently run SambaNova's dataflow architecture because it addresses the memory bottleneck that limits GPU inference. ### The Decode Bottleneck When you send a prompt to an LLM, two distinct phases occur: **Prefill** processes your entire input prompt in parallel. The model reads all your tokens at once, builds the KV cache, and utilization runs high. This is what GPUs were designed for. **Decode** generates output tokens one at a time. Each token depends on all previous tokens, so the work is sequential. The GPU spends most of its time waiting for data from memory - utilization drops to 20-40%. This is the decode bottleneck - not a software problem you can optimize away, but a fundamental mismatch between GPU architecture and the workload. ### Why Memory Bandwidth Limits GPUs For a dense 70B model: - Compute: 140B floating point operations per token, H100 delivers 2,000 TFLOPS = ~0.07ms - Memory: 70B parameters × 2 bytes (FP16) = 140GB, H100 bandwidth 3.35 TB/s = ~42ms The GPU is compute-bound for 0.07ms and memory-bound for 42ms per token. A 600x difference. ### How Dataflow Architecture Changes the Equation SambaNova's RDU (Reconfigurable Dataflow Unit) solves this with: 1. **Assembly Line vs Job Shop**: Data flows through stationary workstations instead of workers walking to memory 2. **Three Levels of Storage**: SRAM (on-chip), HBM (1TB/rack), DDR (12TB/rack) - software-controlled placement eliminates cache misses 3. **Static Scheduling**: Pre-planned execution with no runtime coordination overhead 4. **Continuous Pipeline**: Operations fused together, intermediate results stay on-chip ### Real-World Numbers On Infercom's EU infrastructure: - gpt-oss-120b: 713 tok/s (up to 772 on short prompts), 388ms TTFT on 10K input - MiniMax M2.7 Ultraspeed: 428 tok/s (up to 444 on short prompts), 690ms TTFT on 10K input GPU-based providers typically achieve 50-150 tok/s on large models. The 3-10x speed difference reflects the architectural advantage. ### Energy Efficiency Infercom's SambaNova racks draw about 10 kW each. Equivalent GPU infrastructure runs 40-50 kW per rack. Stanford's Hazy Research measures dataflow delivering up to 5x more "intelligence per watt" than GPU-based inference. ### What's Next: Disaggregated Inference The industry is moving toward disaggregated inference - GPUs handle compute-bound prefill, specialized chips handle memory-bound decode. NVIDIA announced this direction at GTC 2026, and SambaNova partnered with Intel on a similar architecture. URL: [https://infercom.ai/blog/how-dataflow-delivers-speed](https://infercom.ai/blog/how-dataflow-delivers-speed) --- ## Blog: LLM Inference Speed Explained ### LLM Inference Speed Explained: TTFT, Throughput, and What Actually Matters Every inference provider talks about speed. Marketing pages show impressive numbers. But what does "fast" actually mean? Two problems make this hard to parse. First, when one provider advertises 400 tokens per second and another boasts sub-200ms latency, they're measuring completely different things. Both claims can be true - and neither might matter for your use case. Second, "fast" is always relative - but relative to what? When a GPU-based provider claims to be fast, they usually mean fast compared to other GPU providers. That's a reasonable baseline within their category, but it tells you nothing about absolute performance. A provider running on specialized inference hardware might deliver 5-10x the speed, making the GPU comparison irrelevant. ### The Three Metrics That Define Speed When you send a request to an LLM API, the response doesn't arrive all at once. It flows in stages: **Full latency breakdown for a European developer calling a US-based provider:** - Network round-trip: 80-150ms (speed of light through fiber) - TLS handshake: 1-3 additional round trips on new connections - Gateway overhead: 10-50ms (authentication, rate limiting, routing) - Queue time: 0ms to seconds (depends on load and capacity) - Prefill: varies by prompt length - Decode: varies by output length and throughput **Prefill vs Decode:** - Prefill processes your entire input prompt in parallel. This phase is compute-bound. - Decode generates output tokens one at a time, sequentially. This phase is memory-bound. GPUs excel at prefill but struggle with decode (utilization drops to 20-40%). Specialized inference chips like SambaNova's RDU, Groq's LPU, and Cerebras WSE are architected for the memory-bound decode phase, delivering 3-10x faster decode on large models. ### Time to First Token (TTFT) TTFT measures the latency from when you send a request to when the first token starts streaming back. **Thresholds:** - Under 300ms: Excellent for interactive chat - 300-600ms: Acceptable for most applications - Over 600ms: Users start noticing the delay - Over 1 second: Users assume something is wrong TTFT scales with input length. "300ms TTFT" on a short prompt is unremarkable. "300ms TTFT" on 10K tokens is impressive. **Infercom EU infrastructure (10K input tokens):** - gpt-oss-120b: 388ms TTFT - MiniMax-M2.5: 619ms TTFT TTFT matters most for interactive applications: chat interfaces, coding assistants in IDEs, voice AI. ### Output Throughput (Tokens per Second) Throughput measures how fast tokens stream after the first one appears. For a 1,000-token response, the difference between 100 tok/s and 400 tok/s is 10 seconds versus 2.5 seconds. The only fair comparison is the same model across different providers: - DeepSeek R1 671B: 30-80 tok/s (GPU) vs 250+ tok/s (SambaNova RDU) - Llama 3.3 70B: 50-150 tok/s (GPU) vs 2,100 tok/s (Cerebras WSE) **Infercom EU infrastructure:** - gpt-oss-120b: 713 tok/s (up to 772 on shorter prompts) - MiniMax M2.7 Ultraspeed: 428 tok/s (up to 444 on shorter prompts) Throughput matters most for agentic workflows, batch processing, and long-form content generation. ### End-to-End Latency End-to-end latency = TTFT + (output tokens / throughput) For a 1,000-token output at 100 tok/s with 500ms TTFT: 10.5 seconds total At 400 tok/s with 600ms TTFT: 3.1 seconds total **Infercom EU infrastructure (10K input, 1K output):** - gpt-oss-120b: 1.789 seconds end-to-end - MiniMax-M2.5: 3.103 seconds End-to-end latency matters for batch processing, backend API calls, and SLA-bound applications. ### Why These Metrics Conflict The prefill and decode phases compete for hardware resources. Providers optimize for either TTFT or throughput, not both. This tradeoff is changing with disaggregated inference - running prefill and decode on separate hardware. At GTC 2026, NVIDIA announced this direction, and SambaNova partnered with Intel on a heterogeneous architecture. **Context length compounds everything:** - 1,000 tokens: 1 million attention calculations - 10,000 tokens: 100 million attention calculations - 128K context prompt: ~4 seconds to prefill - 1M context prompt: ~77 seconds to prefill ### Matching Metrics to Use Cases - **Interactive chat**: Prioritize TTFT - **Agentic workflows**: Prioritize throughput (agents wait for complete responses) - **Voice AI**: Both matter - **Batch processing**: Throughput and total completion time - **RAG pipelines**: Balance responsiveness with retrieval latency **The Agentic Use Case:** A single coding task might require 50-200+ LLM calls. At 100 tok/s, a 300K token session takes ~50 minutes. At 400 tok/s: ~12 minutes. That's 38 minutes saved per task. ### Consistency Under Load Questions to ask providers: - P50 vs P99 latency: If P99 is 5x higher than P50, you'll have frustrated users - Rate limits: Can you actually hit advertised speeds? - Performance degradation at scale: Does throughput hold at 100 requests/second? ### The Bottom Line 1. Speed claims without context are meaningless. Always ask: which metric, what input length, what baseline? 2. Your workload determines which metric matters. 3. Consistency under load matters as much as peak performance. 4. Network latency is physics. For EU users, Munich is always faster than Virginia. ### Why We Call It "Ultraspeed" MiniMax M2.7 Ultraspeed on Infercom: 1. SambaNova dataflow architecture - purpose-built inference hardware 2. 428 tokens per second measured throughput 3. 690ms TTFT on 10K input tokens, below 150ms on short prompts (from Germany) Test it yourself: [https://benchmark.infercom.ai](https://benchmark.infercom.ai) Published benchmarks: [https://infercom.ai/performance](https://infercom.ai/performance)