Mistral Releases Mistral Large 3 with 675B Parameters and Three Ministral 3 Models Under Apache 2.0
Mistral AI has released Mistral 3, consisting of Mistral Large 3—a sparse mixture-of-experts model with 675B total parameters and 41B active parameters—and three Ministral 3 models at 3B, 8B, and 14B parameters. All models are released under the Apache 2.0 license with multimodal capabilities including image understanding.
Mistral Large 3 — Quick Specs
Mistral Releases Mistral Large 3 with 675B Parameters and Three Ministral 3 Models Under Apache 2.0
Mistral AI has released Mistral 3, consisting of Mistral Large 3—a sparse mixture-of-experts model with 675B total parameters and 41B active parameters—and three Ministral 3 models at 3B, 8B, and 14B parameters. All models are released under the Apache 2.0 license with multimodal capabilities including image understanding.
Mistral Large 3 Technical Specifications
Mistral Large 3 is a sparse mixture-of-experts architecture trained from scratch on 3,000 NVIDIA H200 GPUs. The model uses 41B active parameters and 675B total parameters, making it Mistral's first MoE model since the Mixtral series.
According to Mistral AI, the model ranks #2 in the open-source non-reasoning models category on the LMArena leaderboard (#6 among all open-source models overall). The company claims the instruction-tuned version achieves parity with the best instruction-tuned open-weight models on general prompts while demonstrating what it calls "best-in-class performance" on multilingual conversations in languages other than English and Chinese.
Both base and instruction fine-tuned versions are available under Apache 2.0. A reasoning variant is announced as coming soon.
Ministral 3 Series Details
The Ministral 3 series includes three model sizes: 3B, 8B, and 14B parameters. For each size, Mistral releases base, instruct, and reasoning variants—all with multimodal image understanding capabilities under Apache 2.0.
Mistral AI claims the Ministral 3 reasoning 14B variant achieves 85% accuracy on AIME 2025. The company states that instruct models "match or exceed the performance of comparable models while often producing an order of magnitude fewer tokens."
Infrastructure and Deployment
All Mistral 3 models were trained on NVIDIA Hopper GPUs with HBM3e memory. Mistral collaborated with NVIDIA, vLLM, and Red Hat to optimize deployment:
- Mistral Large 3 can run on a single 8×A100 or 8×H100 node using vLLM
- A checkpoint in NVFP4 format built with llm-compressor is available
- NVIDIA integrated Blackwell attention and MoE kernels for the sparse architecture
- Support for prefill/decode disaggregated serving and speculative decoding on GB200 NVL72
- Ministral models optimized for NVIDIA DGX Spark, RTX PCs, and Jetson edge devices
Inference support is enabled through TensorRT-LLM and SGLang for the complete model family.
Availability
Mistral 3 is available immediately on Mistral AI Studio, Amazon Bedrock, Azure Foundry, Hugging Face, Modal, IBM WatsonX, OpenRouter, Fireworks, Unsloth AI, and Together AI. NVIDIA NIM and AWS SageMaker availability is listed as coming soon.
Pricing information has not been disclosed. Model documentation and research papers are available through Mistral AI's documentation hub and Hugging Face.
What This Means
Mistral Large 3's 675B parameter count with 41B active parameters positions it as one of the largest openly-licensed MoE models available. The Apache 2.0 license removes commercial restrictions that limit other "open" models. The simultaneous release of smaller Ministral variants (3B-14B) with reasoning capabilities addresses the growing demand for edge deployment and cost-efficient inference, though independent verification of Mistral's performance claims on multilingual tasks and token efficiency will be necessary to confirm competitive positioning.
Related Articles
Mistral AI Releases Shieldstral-1.0-3B, a 3B-Parameter Policy-Adaptive Safety Classifier
Mistral AI has released Shieldstral-1.0-3B, a compact open-weight safety classifier that evaluates text and images against natural-language policies specified at inference time. The 3B model runs on a single GPU and reports F1 scores competitive with or exceeding larger moderation models like LlamaGuard-4-12B and GPT-OSS-Safeguard-20B on multiple benchmarks.
NVIDIA Releases Nemotron VoiceChat 11B, an Open Full-Duplex Speech Model with Live Tool Calling
NVIDIA has released NemotronLabs VoiceChat 11B, an 11-billion-parameter end-to-end full-duplex speech model that unifies streaming speech understanding and generation in one architecture. The model claims to be the first open full-duplex system to support live tool calling during natural conversation, with ~450ms turn-taking latency.
Mistral Releases Shieldstral, a 3B Open-Weights Safety Classifier That Matches Models 7x Its Size
Mistral has released Shieldstral, a 3B open-weights safety classifier that reframes content moderation as a policy-adaptive question-answering task. The model claims to match or outperform guard models up to 7x its size on text safety and multimodal benchmarks, and runs on a single 16GB GPU.
LG AI Research Releases K-EXAONE 2.0, a 750B-Parameter Open-Weight MoE Model with 262K Context
LG AI Research has released K-EXAONE 2.0, a 750-billion-parameter mixture-of-experts language model with 37B active parameters, a 262,144-token context window, and support for 10 languages. The model is open-weighted under Apache 2.0 and claims competitive results against Qwen3.5, GLM-5.1, and DeepSeek-V4 Pro on reasoning, coding, and long-context benchmarks.
Comments
Loading...