Sovereign AI models hosted in Europe

GLM-5.3 by Z.ai is a 744B open-weight coding and agentic model, post-trained on the GLM-5.2 base with a 1M-token context and low, high, and max reasoning effort levels.

Kimi K3 by Moonshot: open-weight 2.8 trillion parameter MoE with 104 billion active, native vision, and a 1M token context for long-horizon coding and agentic work.

GLM-5.3-Flash by Z.ai is a natively multimodal 320B MoE with 18B active parameters, hybrid sparse and linear attention, a 1M-token context, and MIT open weights for coding and agent work.

Qwen3.8-2.4T-A95B is the open-weight release of Alibaba's Qwen3.8-Max flagship, a 2.4 trillion parameter Mixture-of-Experts activating 95B per token.

Qwen3.8 Flash-Next by Alibaba is a multimodal MoE with a 125B backbone and 6B active parameters, an early look at the Qwen4 architecture with 262K native context for coding and agents.

DeepSeek V4.1 Flash is a 552B multimodal MoE with a causal encoder-decoder design, 8B to 16B active parameters, a 1M-token context, and MIT weights for fast agentic coding.

DeepSeek V4 Pro, flagship 1.6T MoE with 1M context and hybrid thinking for frontier reasoning and agents.

DeepSeek V4 Flash 0731, the July 2026 retrained checkpoint with major agentic and coding gains over the original release.

DeepSeek V4 Flash, lightweight 284B MoE with 1M context and hybrid thinking for cost-efficient agents.

Qwen3.8 27B by Alibaba is a dense 27B open-weight vision-language model with image and video input, a 262K native context, thinking on by default, and Apache 2.0 weights.

GLM-5.2 by Z.ai is a 744B open-weight (MIT) mixture-of-experts coding model with a 1M-token context, IndexShare sparse attention, and selectable thinking-effort levels.

MiniMax M3: natively multimodal 428B Mixture-of-Experts model with ~23B active parameters, MiniMax Sparse Attention for long context, and 59.0% on SWE-Bench Pro

Kimi K2.6 by Moonshot: open-weight agentic model with long-horizon coding, 300-agent swarms, video understanding, and 256K context.

Kimi K2.7 Code by Moonshot: open-weight, code-specialized agentic model on a 1 trillion parameter MoE, tuned for long-horizon software engineering and tool use.

MiniMax M2.5: 229B MoE model trained on hundreds of thousands of real environments, scoring 80.2% on SWE-Bench Verified and 76.3% on BrowseComp

Qwen3.5 397B A17B by Alibaba: large sparse MoE model with 397B total (17B active) parameters for advanced reasoning and coding.

Qwen3.6-27B by Alibaba: 27B open-weight multimodal model with strong agentic coding (77.2 SWE-bench Verified) and reasoning across 262K context.

Gemma 4 31B is Google DeepMind's 30.7B parameter dense open model, the largest of the Gemma 4 family, with 256K context, thinking mode, and native function calling.

Qwen3.6 35B-A3B by Alibaba: sparse MoE (35B total, 3B active) with vision, reasoning, and agentic coding for repository-level tasks.

Qwen3.6 35B A3B FP8 by Alibaba: fine-grained FP8 quantized variant of the sparse MoE model with near-identical performance.

Qwen3.5-122B-A10B by Alibaba: sparse MoE multimodal model with 122B total (10B active) parameters and 262K context for reasoning, vision, and agentic tasks.

GLM 4.7 Flash by Zhipu: 30B MoE (3B active) with 200K context, tuned for local deployment and fast agentic coding.

Qwen3.5 9B by Alibaba: 9B multimodal model with vision, tool calling, and reasoning for balanced performance and efficiency.

Qwen3 235B A22B Instruct 2507 by Alibaba, an instruction-tuned MoE model with a 262K context window.

GPT-OSS 120B by OpenAI: open-weight Apache 2.0 reasoning model, 117B params (5.1B active), near o4-mini on reasoning, tools and structured output.

Mistral Small 4 (March 2026) unifies reasoning, coding, and multimodal in one MoE model with 256K context and configurable reasoning effort.

GPT-OSS 20B by OpenAI: open-weight Apache 2.0 model, 21B params (3.6B active), near o3-mini, for efficient and on-device reasoning.

Qwen3 Coder Next by Alibaba, an 80B ultra-sparse MoE model with 3B active parameters and 256K context for efficient local coding agents.

Qwen3 VL 30B A3B Instruct by Alibaba, an efficient open-weight MoE vision-language model with 3B active parameters and 131K context.

Llama 3.3 70B Instruct: refined 70B model with improved multilingual performance, structured output, and tool calling for reasoning workloads.

Llama 3.3 70B Instruct in an FP8 build, Meta's multilingual open-weight workhorse with 128K context, quantized for faster serving at near-identical quality.

Llama 3.1 8B Instruct: efficient open-weight small model with a 131K context, tool calling, and structured output for production deployments.

OpenAI's 1.55B-parameter speech-to-text model supporting 99 languages with 10-20% error reduction over Whisper Large v2, released November 2023.

OpenAI's pruned 809M-parameter speech-to-text model with 4 decoder layers, delivering much faster multilingual transcription at minimal quality loss.

Google Gemma 4 26B-A4B IT, mixture-of-experts open model with only 3.8B active parameters, multimodal reasoning, and function calling.

Google Gemma 4 31B IT, instruction-tuned 31B dense model with multimodal reasoning, function calling, and 256K context.

Mistral OCR 4.1, the current model behind Mistral's Document AI stack, adding paragraph-level bounding boxes and block-level confidence scores to extracted text.

Mistral Medium 3.5 from Mistral — text model on the Opper gateway.

Mistral Medium 3.5, dense 128B open-weight model with 256k context optimized for agentic and coding workflows.

Mistral Small 3.2 24B Instruct is a 128K-context open-weight text and tool-calling model tuned for reliable instruction following and lean deployments.

Mistral Small 3.1 24B is an open-weight instruction-tuned model with 128K context, vision, and function calling at around 150 tokens per second.

Mistral Ministral 3 14B: largest Ministral 3 model, 256k context, Apache 2.0, tuned for local and edge deployment.

Mistral Ministral 3 3B: smallest Ministral 3 model, 256k context, Apache 2.0, ultra-efficient edge deployment.

Mistral Ministral 3 8B, edge-optimized multimodal model with 256k context, vision, and structured outputs.

Mistral Large 3, open-weight mixture-of-experts model with 256k context and native multimodal vision.

Mistral Codestral 25.08: code-optimized model for fill-in-the-middle completion and fast code generation, 128k context.

Voxtral Mini Latest routes to the current Voxtral Mini, a 32K-context speech model for voice agents with native audio understanding.

Voxtral Mini Transcribe Realtime 2602 delivers streaming live transcription with 13-language support, speaker diarization, and a 4B edge-deployable model.

Voxtral Mini Transcribe Realtime Latest routes to the current streaming transcription model, a 4B 13-language model for voice agents and live apps.

Voxtral Mini Transcribe Realtime by Mistral: streaming speech-to-text with configurable delay for live transcription, voice agents, and real-time subtitling.

Voxtral Small 2507 by Mistral: 24B audio-native LLM with function calling and structured outputs for speech understanding and agentic chat.

Voxtral Small (latest) by Mistral: open-weight 24B audio-input instruct model that tracks the newest snapshot, with function calling and structured outputs.

Mistral OCR 3, document understanding model with a 74% win rate over OCR 2 on forms, tables, and handwritten text.

Voxtral Mini 2602 is a 4B streaming speech-to-text model for live applications, with speaker diarization and 13-language support.

Mistral Codestral Embed: embedding model specialized for semantic code search and retrieval, 8k context.

Mistral Codestral Embed 25.05: code retrieval embedding model Mistral reports beating Voyage Code 3, Cohere, and OpenAI.

Mistral Embed, general-purpose text embedding model producing 1024-dimensional vectors for semantic search and retrieval.

Mistral Embed 2312, efficient text embedding model with an 8k context for RAG and semantic search applications.

Voxtral Mini TTS by Mistral: 4B text-to-speech with zero-shot voice cloning from about 3 seconds of reference audio, supporting 9 languages.

Voxtral Mini TTS (latest) by Mistral: text-to-speech with zero-shot voice cloning and multilingual support, tracking the newest model snapshot.

Qwen3 14B, Alibaba's April 2025 dense model with switchable thinking and direct modes for reasoning and tool use at 40K context.

Qwen3-Embedding 8B by Alibaba, a 4096-dimensional embedding model with 36 layers and 32K context for dense retrieval and retrieval-augmented generation.

Qwen3 30B A3B by Alibaba, an efficient MoE model with 3B active parameters and switchable thinking and non-thinking modes.

Qwen3 32B by Alibaba, a dense 32B model with switchable thinking and non-thinking modes for reasoning and speed.

Qwen3 8B dense model on Fireworks

Multilingual E5 Large by Microsoft: semantic search embedding model supporting 100 languages with 1024-dimensional representations.

Multilingual E5 Large Instruct by Microsoft: instruction-tuned embedding model with 1024 dimensions for multilingual semantic retrieval across 100 languages.

BGE-reranker-v2-m3 by BAAI, lightweight 600M cross-encoder reranker built on BGE-m3 for fast multilingual relevance scoring.

Talkie 1930 is a 13B open-weight model trained only on pre-1931 English text, giving a 1930 knowledge cutoff for historical reasoning and contamination-free research.

IBM Docling: open-source OCR and document parsing with layout analysis, handling many formats including PDF, images, Office documents, and audio.

Apertus 70B, the Swiss AI Initiative's fully open model from EPFL and ETH Zurich, trained on 15T tokens with over 1,800 natively supported languages.

Brick Complexity Pro, Regolo's hosted prompt-complexity classifier that grades queries easy, medium, or hard so routing picks the right model tier.

green-l-raw GreenPT Backed by Mistral Small 3.2 24B Direct access to the same GreenPT-backed model as green-l without the built-in system prompt. Input €0.25 Output €0.80 Context 128k Max output 32k Released Jun 2025 Text Images Documents Multilingual No system prompt

green-r-raw GreenPT Backed by GPT-OSS Direct access to the same reasoning stack as green-r without the GreenPT system prompt. Input €0.35 Output €0.95 Context 128k Max output 32k Released Aug 2025 Text Images Documents Multilingual No system prompt

green-l GreenPT Backed by Mistral Small 3.2 24B GreenPT-branded chat model tuned for multilingual writing, image understanding, and Dutch grammar guardrails. Input €0.25 Output €0.80 Context 128k Max output 32k Released Jun 2025 Text Images Documents Multilingual Writing assistant Dutch grammar guardrails

green-r GreenPT Backed by GPT-OSS GreenPT-branded reasoning model for advanced analysis, writing, and content generation. Input €0.35 Output €0.95 Context 128k Max output 32k Released Aug 2025 Text Images Documents Multilingual Advanced reasoning Writing & content generation

Expressive multilingual speech synthesis across 50+ languages, with voice selection by reference id, multi-speaker dialogue and prosody control.

Fish Speech S2.1 Pro Realtime from Fish Audio — text model on the Opper gateway.

Faster Whisper Large V3 is a multilingual speech-to-text model supporting 99 languages, a CTranslate2 reimplementation of OpenAI's Whisper that runs up to 4x faster at the same accuracy.

Sao10K's L3.3 70B Euryale v2.3, a 70B Llama 3.3 creative roleplay model and the direct successor to v2.2.

ALIA 40B Instruct 2601, Spain's publicly funded 40B multilingual model from the Barcelona Supercomputing Center, Apache 2.0 and pretrained across 35 European languages.

Clef decision model by Cloudflare (27B), hosted by Opper in the EU. Answers noul (yes/no), choice and score questions with a probability for every option instead of text. Text or JSON state up to 16,384 tokens including the questions. Use POST /v3/compat/v1/systemone with model opper/clef. Weights: https://huggingface.co/Cloudflare/clef.

Clef-flash decision model by Cloudflare (9B), hosted by Opper in the EU. Answers noul (yes/no), choice and score questions with a probability for every option instead of text. Text or JSON state up to 16,384 tokens including the questions. Use POST /v3/compat/v1/systemone with model opper/clef-flash. Weights: https://huggingface.co/Cloudflare/clef-flash.

Kev-4B System One decision model, a fine-tune by Jared Palmer of Alibaba's Qwen3.5-4B-Base, for typed yes/no decisions (noul), classification (choice) and rubric scoring (score), returning a probability for every option instead of text. Text or structured text input only, up to 8,192 tokens for the state plus one question; trained on states of up to 384 tokens, so accuracy on long documents is lower. Use POST /v3/compat/v1/systemone with model opper/kev-4b. Weights by Jared Palmer: https://huggingface.co/jaredpalmer/kev-4b. Base model: Alibaba's Qwen3.5-4B-Base, https://huggingface.co/Qwen/Qwen3.5-4B-Base.

Undi95's ReMM SLERP L2 13B, a Llama 2 SLERP merge recreating the MythoMax recipe with updated component models.

Laya System One decision model by ConvAI Innovations (Nandha Kishor M), a non-autoregressive ModernBERT-large encoder (421M) for typed yes/no decisions (noul), classification (choice) and rubric scoring (score), returning a probability for every option instead of text. Text or structured text input only; the English checkpoint reads 512 tokens and truncates longer input. Use POST /v3/compat/v1/systemone with model berget/convaiinnovations/laya. Weights: https://huggingface.co/convaiinnovations/laya (Apache-2.0).

KBLab's KB Whisper Large is a 2-billion-parameter Swedish speech-to-text model trained on 50,000 hours of audio, cutting word error rate by an average 47% versus Whisper-large-v3.

Norwegian and English speech-to-text from the National Library of Norway, a 1.55B-parameter Whisper model trained on 66,000 hours including parliamentary speeches and broadcasts.

Pianissimo from Klang — speech to text model on the Opper gateway.

roc, Evroc's enterprise AI agent for writing, data analysis, coding support, and knowledge retrieval over company documents and systems.