Mistral Large 3
Mistral Large 3 is an open-weight, sparse mixture-of-experts model from Mistral AI, released on December 2, 2025 under Apache 2.0, with 41B active and 675B total parameters. It accepts text and image input.
Superseded. Mistral’s model list now also shows Mistral Large 4 (v26.10). Mistral Large 3 remains generally available.
Released
December 2, 2025
Type
multimodal
License
open-weight
Apache 2.0
Context
256,000 tokens
Pricing
Input
$0.50 / 1M tokens
Output
$1.50 / 1M tokens
List price on the developer's own pages, last checked , more than 7 days ago: it may have changed since. API model mistral-large-2512.
Capabilities
Architecture
Parameters: 675B total, 41B active (mixture of experts)
Sources
Every figure on this page was checked against these pages on . A figure we could not source is left out.
- Mistral AI news: Mistral 3 (opens in a new tab)
- Mistral docs: Mistral Large 3 (opens in a new tab)
- Mistral docs: Models overview (opens in a new tab)
Spotted something out of date? Tell us.
Links
In the news: Mistral Large 3 and Mistral AI
MarkTechPost
Mistral AI Releases Voxtral 3: Speech-to-Text, TTS and a Voice Agent SDK
Mistral AI has released Voxtral 3, a family of speech models and a voice agent SDK. The release includes three transcription models, one text-to-speech model, and an SDK integrated with LiveKit, designed for listening, reasoning, tool calls, and speaking. The Voxtral 3 family comprises Voxtral Small Transcribe 3, the flagship batch transcription and diarization model, which also supports translation, speech Q&A, summarization, and function calling. Voxtral Mini Transcribe 3 is a 4B parameter model for more economical batch jobs, while Voxtral Mini Transcribe Realtime 3 offers streaming ASR with end-of-turn detection. The Voxtral Speech 3 model, also referred to as Voxtral Mini TTS 3, handles text-to-speech. All transcription models feature diarization and context biasing capabilities, with batch transcription supporting 50 languages, including 24 official EU languages. Voxtral Mini Transcribe 3 has 4B parameters and a 32k context window. The Voxtral voice agent SDK operates as a three-stage streaming pipeline. It utilizes Realtime Transcribe 3 for listening and turn detection, a general LLM like Mistral Small 4 for reasoning and tool execution, and Voxtral Speech 3 for replies that match the conversation's tone. Each turn is completed in under one second. Performance benchmarks show Voxtral 3 models ranking highly on independent speech-to-text evaluations. Small Transcribe 3 achieved a 2.1% AA-WER, placing it third on Artificial Analysis. In real-time performance, it recorded a 7.4% WER on FLEURS 35 at 240 ms, compared to Gemini 3.5 Realtime's 4.2% WER.
MarkTechPost
Mistral AI Releases Timestral: A 209M Forecasting Model Family That Outperforms Google’s TimesFM-3
Mistral AI has released Timestral, its first family of time series foundation models. This new set includes four zero-shot forecasting models, with parameter counts ranging from 5 million to 209 million. The largest model in the Timestral family has achieved the top rank among pretrained, zero-shot models on the GIFT-Eval, TIME, and fev-bench benchmarks. Timestral models can perform probabilistic forecasting on standard server CPUs, eliminating the need for GPUs during inference. The models offer open weights and inference code for developers. The largest 209M parameter model requires 837 MB in fp32 and can forecast multivariate targets jointly, accepting past-only, known-future, and categorical covariates. The performance of the 209M Timestral model is notable, with Mistral reporting a fev-bench SQL skill score of 50.39, surpassing Google's TimesFM-3 score of 48.99. On GIFT-Eval, its MASE score of 0.658 leads t0-gamma by a narrow margin of 0.003. Timestral is a causal, patch-based transformer utilizing time attention and variate attention. Unlike classical tools that fit one series at a time or deep models that train from scratch on each dataset, Timestral transfers patterns across domains. Each model outputs 21 quantiles per time step. The 209M model features 24 layers and an embedding dimension of 768, with each patch covering 32 time steps.
MarkTechPost
Mistral AI Releases Mistral Large 4 (Le Chonk): A 1.05T Parameter Multimodal MoE Model
Mistral AI announced the public preview release of Mistral Large 4 (ML4), internally codenamed 'Le Chonk,' on October 6, 2026. This new model is a granular Mixture of Experts (MoE) architecture, boasting a total of 1.05 trillion parameters. Specifically, it features 49 billion active parameters per token, a 1.6 billion parameter vision encoder for native image input, and an extensive 1 million token context window. The model was trained from scratch using 3,800 NVIDIA Grace Blackwell GPUs within Mistral's own European data centers. The API for ML4 is now accessible, with pricing set at $1.36 per 1 million input tokens and $4.18 per 1 million output tokens. However, the model weights are scheduled for release at the end of October, meaning self-hosting capabilities will not be available until then. Mistral AI highlights ML4's standout performance in cybersecurity, reporting a 93% score on the Cybench benchmark and 82% on CyberGym-E2E. The company noted that several closed frontier models achieved near-zero scores on these benchmarks due to task refusal. The architecture of ML4 is described as a hybrid instruct-and-reasoning MoE that natively accepts image inputs. The efficient activation of only approximately 4.7% of the total weights per token allows a model of this scale to be served at mid-tier pricing. While the full 1.05 trillion parameters require memory, the activation count dictates the compute resources needed, not the overall hardware expenditure for users. Mistral has not yet disclosed the specific expert count, top-k routing mechanisms, or layer layout, details that are expected to be provided upon the release of the model weights. The training data for ML4 encompassed over 160 languages, including all official European Union languages, indicating a broad linguistic capability. Mistral AI, founded in 2023 by former Google DeepMind and Meta AI researchers, is a European artificial intelligence company focused on developing open and efficient large language models. The company has previously released models like Mistral 7B and Mixtral 8x7B, which gained traction for their performance and accessibility. The development of ML4 signifies a significant step forward in the company's ambition to compete with larger, established AI labs by pushing the boundaries of model scale and multimodal capabilities while maintaining a focus on efficiency and European data sovereignty. The release of ML4 positions Mistral AI as a key player in the competitive landscape of advanced AI development, particularly with its emphasis on specialized benchmarks like cybersecurity.
MarkTechPost
Cohere Releases North Small Translate: A 218B MoE Translation Model That Scores 83.6 on WMT26 Across 50 Languages
Cohere has released North Small Translate, an open-weight machine translation model developed by Cohere and Cohere Labs. This model is a sparse Mixture-of-Experts (MoE) architecture, featuring a total of 218 billion parameters, with 25 billion active parameters per token. North Small Translate supports translation across 50 languages, ranging from Albanian to Vietnamese. In evaluations conducted by Cohere on the WMT26 benchmark, the model achieved an average score of 83.6 across all supported languages. Cohere claims this performance surpasses that of commercial translation services like DeepL and Google Translate, as well as other open-source models such as GLM 5.2 and Mistral Large 3. The model is accessible through various deployment options: it can be used freely via Cohere's API up to specified rate limits, self-hosted for non-commercial use, or licensed for commercial applications. The development of North Small Translate represents a return to the foundational problem of machine translation, which was a key focus of the original Transformer architecture introduced by Google researchers in 2017 with the paper "Attention Is All You Need." The Transformer's initial breakthroughs were demonstrated on WMT 2014 English-to-German and English-to-French translation tasks. Cohere frames the release of this dedicated translation model as a matter of digital sovereignty, asserting that organizations unable to communicate globally may struggle to maintain their sovereignty. North Small Translate is the inaugural model within Cohere's "North" family of models and follows previous multilingual efforts such as Tiny Aya and Command A Translate. The model's real-world translation quality was shaped through collaboration with RWS, an organization whose Language Weaver scientists and language experts contributed to its development. Architecturally, North Small Translate is a decoder-only sparse MoE Transformer. It incorporates 128 experts, with 8 experts activated per token, alongside shared experts applied universally. The routing mechanism employs a sigmoid function over expert logits, normalized across the selected top-k experts. Its attention mechanism includes interleaved sliding-window layers (with a window size of 4096 and RoPE embeddings) and global layers that do not utilize positional embeddings, in a 3:1 ratio. This specific attention configuration was previously introduced in Cohere's Command A model. The model supports an input and output context window of 16,000 tokens, exclusively for text. It underwent post-training specifically to enhance translation quality. Approximately 11.5% of the model's weights are active for each token processed, and the per-token compute load corresponds to the 25 billion active parameters, although the full 218 billion parameters must be held in memory.
MarkTechPost
Mistral AI Releases Shieldstral 1.0 3B: An Open-Weights Policy-Adaptive Multimodal Safety Classifier Matching Models 7× Its Size
Mistral AI has released Shieldstral 1.0 3B, an open-weights, policy-adaptive multimodal safety classifier designed to handle content moderation by framing it as a single yes/no question rather than relying on a predefined taxonomy of harm categories. Unlike many existing guardrail models that embed their category lists directly into their weights, requiring retraining for new deployment contexts, Shieldstral allows operators to define policies using plain-language questions at inference time. The model then returns a calibrated safety score after a single forward pass. This approach enables dynamic policy adaptation, allowing content that might be acceptable in one context, such as a cybersecurity research tool, to be flagged as harmful in another, like a mental-health platform. Shieldstral 1.0 3B is built upon the Ministral-3-3B-Base-2512 model and incorporates a native Pixtral vision encoder. It is released under the permissive Apache 2.0 license, permitting both commercial and non-commercial use. In performance evaluations, Shieldstral reports an 84.9% average F1 score for text safety, a performance level that matches the GPT-OSS-Safeguard-20B model. For multimodal safety, it achieves an 83.8% F1 score, outperforming all other baselines evaluated by Mistral. The model's architecture and size make it deployable locally. Shieldstral-1.0-3B requires 16GB of VRAM when using BF16 precision and can operate on a single GPU. Serving the model is facilitated through multiple established pathways, including vLLM (version 0.26.0 and later, recommended), llama.cpp via GGUF conversion with Q8_0, Q5_K_M, or Q4_K_M quantization, SGLang, and the Transformers library. Fine-tuning is also supported through the Axolotl framework. The classifier's output is a single token, which contributes to significantly lower latency and cost compared to reasoning-based guardrails like GPT-OSS-Safeguard-20B. The 16GB VRAM requirement positions Shieldstral 1.0 3B as an accessible solution for early-stage AI product teams that may not have the resources to contract with dedicated moderation vendors. Furthermore, its open license and self-hosting capabilities are advantageous for mid-market and enterprise organizations that require guardrails within their Virtual Private Cloud (VPC) or on-premises infrastructure to comply with data residency regulations and facilitate auditing. Multi-tenant SaaS providers can also benefit, as a single checkpoint can be configured to enforce distinct policies for each customer. The model is particularly relevant for industries such as consumer social media and user-generated content platforms, educational technology and child safety applications, healthcare and mental health services, fintech and insurance support automation, gaming and voice chat services, and marketing technology.