Mistral-Large-4-1T-A52B-NVFP4-GGUF Vision-Language Model: 8GB+ VRAM

Mistral-Large-4-1T-A52B-NVFP4-GGUF Vision-Language Model: 8GB+ VRAM

At a Glance

Item Value
Repository ggml-org/Mistral-Large-4-1T-A52B-NVFP4-GGUF
Family guide Mistral-Large-4-1T guide (1 articles)
Publisher guide Mistral AI: models and licenses
Published 2026-10-29
License mistral-ecosystem-open-weight-license
Formats GGUF
Source type Primary source (the publisher itself)

Values determined by this site’s code when the information was collected. Dates are JST.

Overview

ggml-org/Mistral-Large-4-1T-A52B-NVFP4-GGUF has been released by ggml-org, providing an NVFP4 quantized version of Mistral AI’s latest flagship model in GGUF format.

This model is a multimodal image-text-to-text model that supports both text and image inputs. The source model is mistralai/Mistral-Large-4-1T-A52B-NVFP4, which is an NVFP4 quantized checkpoint of Mistral Large 4 trained from scratch using 3,800 NVIDIA Grace Blackwell GPUs in Mistral AI’s European data center.

In the GGUF conversion process, the NVFP4 routed experts of the original model are preserved as-is without requantization. Additionally, FP8 layers such as the Attention layers, indexer query projections, shared experts, and the first Dense layer are stored losslessly in BF16, just like the layers that were originally BF16.

Specifications

  • Parameters: Total parameters of 1 trillion (1T), active parameters of approximately 52B (49B per token, 52B including embedding and output layers)
  • Architecture: Mixture-of-Experts (MoE, 128 routed experts, 4 active experts per token, and 1 shared expert)
  • Context Length: 1,000,000 (1M) tokens

Performance

Specific benchmark score numbers and comparison tables against third-party organizations or other models are not provided in the model cards for this model or the source model mistralai/Mistral-Large-4-1T-A52B-NVFP4.

According to an explanation from the publisher Mistral AI, it is reported that NVFP4 format checkpoints achieve significantly reduced effective memory footprints compared to FP8 format checkpoints while delivering performance comparable to FP8 and faster inference processing. It specifically targets leveraging dedicated hardware features available in the NVIDIA Blackwell architecture (such as B200 / B300 / GB200 / GB300).

For response generation, the depth of thought can be controlled via the reasoning_effort parameter. Three setting values are available: 'low', 'medium', and 'high', with the default setting being 'high'. Note that this thinking process feature itself cannot be completely disabled (turned off). Using 'high' is recommended for complex prompts and agent use cases, while 'low' is recommended for quick chat responses.

Additionally, to further accelerate inference, a DSpark speculative decoding draft head consisting of 3 Multi-head Latent Attention (MLA) layers is integrated. This draft head is specified to propose 3 tokens at once from the target model’s hidden states using a 2,048-token sliding window.

As for sampling settings, GGUF metadata records a temperature of 1.0 and top_p of 0.95 as the officially recommended configurations.

Strengths and Use Cases

Mistral Large 4, the base model for this release, is designed as a general-purpose multimodal model capable of handling diverse advanced tasks. The strengths and intended use cases of the original model are as follows:

Multimodal Image and Text Understanding

Supporting image inputs in addition to text inputs, it can be applied to image-text-to-text tasks such as image-based question answering and analysis of contexts where images and text are mixed. The output is in text format.

Large-Scale Long Context and Broad Multilingual Support

It supports an extremely wide context window of up to 1 million (1M) tokens, enabling the analysis of large documents and processing while maintaining long conversation histories. Furthermore, supported languages include Japanese (ja), English (en), French (fr), German (de), Spanish (es), Portuguese (pt), Italian (it), Korean (ko), Chinese (zh), Russian (ru), Arabic (ar), and many others.

Structured Output and Agent/Tool Usage

It supports native function calling (tool use) and strict JSON Schema output. When operating as an agent system, it is recommended to define clear execution environments and use cases in the system prompts, and give instructions to utilize tools effectively. It is also effective to clearly define the set of tools provided and narrow them down to the minimum necessary for the use case.

Task-Specific Reasoning Control (reasoning_effort)

By setting the reasoning_effort parameter, the reasoning level during generation can be adjusted. It is possible to flexibly switch according to the nature of the task, such as selecting the default 'high' for complex prompt interpretation and autonomous agent utilization, or 'low' for quick chat responses.

Hardware Requirements

Estimated requirements (calculated by Local Model Watch)

Your VRAM Quantization File size Est. memory needed
8GB (RTX 4060 / 3060 Ti, etc.) BF16 3.9GB 4.7GB

Memory estimates add a 20% runtime overhead (KV cache, etc.) to the actual size of the distributed files. Actual usage varies with context length, batch size and inference engine. These figures are computed by this site from file sizes, not published by the model’s authors. Compare with other models in our VRAM quick reference. What the quantization names mean: glossary.

Can You Run It Locally?

Runs in Ollama, LM Studio and llama.cpp as-is.

It is distributed in GGUF, so no conversion is needed.

License — mistral-ecosystem-open-weight-license: A custom license from the publisher. Check the original terms directly, including whether commercial use is permitted.

Compiled by this site’s code from the published formats, converted builds we have found, and each engine’s own model registry. “Not found" means we have not seen such a build, not that none exists. License summaries are not legal advice — check the publisher’s original terms before relying on them.

Our Own Measurements

Measurements We Did Not Take

We have not confirmed that this site meets the commercial-use terms of this model’s license (mistral-ecosystem-open-weight-license). Because this site carries advertising, we did not run the model (no CPU run, answers, quantization comparison, conversion or generation).

How to Get It

This model is distributed in GGUF format and runs on compatible tools such as llama.cpp (llama.app).

Distribution Format and Acquisition Command

Provided in GGUF format, an example command for direct download and execution using the server feature of llama.cpp (llama serve) is as follows:

llama serve -hf ggml-org/Mistral-Large-4-1T-A52B-NVFP4-GGUF:NVFP4

Additionally, the source model mistralai/Mistral-Large-4-1T-A52B-NVFP4 is compatible with inference frameworks such as vLLM and SGLang.

License and Conditions

The license for this model is the “Mistral Ecosystem Open-Weight License". This is modified based on the MIT license, and both commercial and non-commercial uses are permitted. However, operating this model as a “Model-as-a-Service" business targeting third parties requires entering into a separate individual commercial license agreement with Mistral AI.

Note that this repository is not a gated model requiring a consent procedure, so it can be downloaded directly on Hugging Face without prior agreement to terms.

Quantized and Converted Variants

→ Scroll horizontally to see all columns

Added Publisher Format Repository Smallest VRAM tier (build, est. memory)
2026-10-29 ggml-org GGUF ggml-org/Mistral-Large-4-1T-A52B-GGUF IQ2_XXS 332.7GB (does not fit a single consumer GPU)
2026-10-30 unsloth GGUF unsloth/Mistral-Large-4-1T-A52B-GGUF Q8_0 2.5GB (fits in 4GB VRAM)

File sizes of each build:

  • Available builds in ggml-org/Mistral-Large-4-1T-A52B-GGUF: IQ2_XXS 277.2GB / Q2_K 354.6GB / Q4_K_M 591.2GB / Q8_0 1038.6GB / BF16 1954.7GB
  • Available builds in unsloth/Mistral-Large-4-1T-A52B-GGUF: Q8_0 2.1GB / BF16 3.9GB / IQ1_S 217.0GB / IQ1_M 241.8GB / IQ2_XXS 281.7GB / Q2_K_XL 352.3GB / IQ3_XXS 394.4GB / Q3_K_XL 458.9GB / Q4_K_XL 563.3GB / Q5_K_XL 729.9GB / Q6_K_XL 829.3GB

This section is appended automatically by Local Model Watch when a converted build of this model appears after publication. Memory figures are estimated from the size of the distributed files. What the quantization names mean: glossary.

Related Articles

What to Read Next

Sources