Mistral Small 4
Mistral Small 4 is a mixture-of-experts model from Mistral AI, released on March 16, 2026, that brings Mistral’s instruct, reasoning (Magistral), multimodal (Pixtral) and agentic coding (Devstral) capabilities into one model, with configurable reasoning effort.
Released
March 16, 2026
Type
multimodal
License
open-weight
Apache 2.0
Context
—
The developer states a context window of “256k” tokens.
Pricing
Input
$0.15 / 1M tokens
Output
$0.60 / 1M tokens
List price on the developer's own pages, checked . Standard tier; cached input $0.015 per 1M tokens.
Capabilities
Architecture
Parameters: 119B total, 6B active per token (8B including embedding and output layers); Mixture-of-Experts with 128 experts, 4 active
Sources
Every figure on this page was checked against these pages on . A figure we could not source is left out.
- Mistral blog: Introducing Mistral Small 4 (2026-03-16) (opens in a new tab)
- Mistral docs: Mistral Small 4 model page (opens in a new tab)
- Mistral docs: API pricing (opens in a new tab)
Spotted something out of date? Tell us.
Links
In the news: Mistral Small 4 and Mistral AI
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
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.
MarkTechPost
Mistral AI Releases Robostral Navigate: An 8B Model Enabling Robots to Navigate Complex Environments Using a Single RGB Camera
Mistral AI released Robostral Navigate this week, marking its first model specifically designed for embodied navigation in robots. This 8 billion parameter model processes RGB images and plain-language instructions to guide a robot through complex environments, including offices, residential buildings, commercial spaces, and outdoor settings. A key feature of Robostral Navigate is its ability to achieve a 76.6% success rate on R2R-CE validation, navigating previously unseen environments using only a single RGB camera, without relying on depth sensors or LiDAR. The model operates by taking a single instruction, such as 'Leave the lobby, walk through the corridor, enter the supply room, and stop to face the second shelf,' and executing the entire task autonomously. This capability is particularly significant as it allows robots to navigate live spaces populated with people and unexpected obstacles, elements not present during its training phase. Traditional navigation systems often require multiple sensors or specialized hardware, whereas Robostral Navigate demonstrates efficiency and effectiveness with standard RGB cameras. Robostral Navigate employs a method called 'pointing' to determine its next movement. Given a task and its observation history, the model predicts the target's image coordinates within the current camera view and the desired orientation upon arrival. This pointing mechanism offers robustness against changes in camera intrinsics and world scale, unlike navigation commands based on metric displacements. When the target is outside the camera's field of view, the model defaults to using displacements within the robot's local coordinate frame, such as 'Move 2 meters forward, 1.5 meters to the left, and turn 25 degrees left.' This dual approach ensures continuous navigation even when visual cues are temporarily unavailable.
MarkTechPost
Mistral AI Releases Leanstral 1.5: An Apache-2.0 Lean 4 Code Agent Model Solving 587 of 672 PutnamBench Problems
Mistral AI released Leanstral 1.5 this week, a code agent model specifically designed for the Lean 4 proof assistant. This new model targets applications in automated theorem proving and proof engineering, with its weights made available under the permissive Apache 2.0 license. A free API endpoint, leanstral-1-5, has also been launched. Leanstral 1.5 represents an update to the earlier Leanstral-2603 model and is part of Mistral AI's Mistral Small 4 family. The model employs a mixture-of-experts (MoE) architecture, activating a subset of specialized sub-networks for each token to maintain computational efficiency while leveraging a large overall capacity. It features 128 experts, with 4 active per token, a total size of 119 billion parameters, and 6.5 billion activated parameters per token. The model supports a context length of 256,000 tokens and can process both text and image inputs, though its output is limited to text. The training process for Leanstral 1.5 involved three stages: mid-training, supervised fine-tuning, and reinforcement learning using CISPO. Two distinct reinforcement learning environments were utilized to shape the model's agentic capabilities. In the multiturn environment, the model was tasked with proving or disproving theorems, iteratively refining its proofs based on feedback from the Lean compiler. The code agent environment immersed Leanstral within a simulated filesystem, enabling it to interact with files, execute bash commands, and utilize the Lean language server for real-time information on goals, errors, and types. This allowed for tasks such as completing partial proofs and creating auxiliary lemmas, with context compaction used to manage long tasks within the context window. Mistral AI reports that Leanstral 1.5 demonstrates strong performance, saturating the miniF2F benchmark and achieving 100% accuracy on its validation set. The model's correctness is verified against target theorems using Mistral's fork of SafeVerify. The model successfully solved 587 out of 672 problems on the PutnamBench, indicating significant advancements in its theorem-proving capabilities.