LLM EU
Menu
Models

Kimi VL A3B Instruct

Moonshot AI

Catalogue only Open weights

Last reviewed:

Sovereignty and hosting

Catalogue only

This model is catalogued for reference. It is not hosted on LLM EU yet, and no first-party endpoint serves it here. Requests to it go to the provider directly, under that provider's own terms.

Summary

The checkpoint is instruction-tuned and ships a chat template. Kimi VL A3B Instruct is a large open-weight vision-language model from Moonshot AI: 16.4B parameters under MIT. It accepts text and images and answers in text. The published configuration records no context window, so this card carries none. The repository was created in April 2025 and is mirrored on Hugging Face. The card is catalogue-only: LLM EU does not run this model. The publisher states no language coverage on the model page.

Specifications

Provider
Moonshot AI (CN)
Context window
Not verified: the provider has not documented this claim.
Parameters
16.4B
Modalities
text, vision
License
mit · licence
Open weights
yes
Release date
09 Apr 2025
Tasks
Vision, Low cost

Languages

Strong: —

Adequate: —

Strong

    Adequate

      Not verified: the provider has not documented this claim. — a language is listed as strong only where a source claims it; the card links the evidence where one exists.

      When to use it

      Strengths

      • open weights are published on Hugging Face and can be downloaded without a vendor agreement
      • mit licence, so commercial self-hosting is permitted without a separate agreement
      • instruction-tuned checkpoint with a chat template in the repository

      Limits

      • catalogue only — LLM EU does not run this model, and nothing on this card implies an LLM EU endpoint
      • no evaluation has been run by LLM EU
      • the published configuration records no context window, so the card carries none
      • at 16.4B parameters, self-hosting needs a single accelerator with enough memory for the weights
      • vision checkpoints are heavier to serve than their parameter count alone suggests, because the image encoder and pre-processing run alongside the language model