Product
Kimi K3: fiche technique, prix API et données
Updated
Moonshot AIAI Models & LLMs
- Context window
- 1,048,576 tokens
- Open weights (downloadable)
- Yes
Data sheet · Kimi K3
Every figure is quoted from the vendor's official page, with the date we read it. If a row is missing, the source does not state it and we do not estimate it.
Memory and answer length
- Context window
- 1,048,576 tokensRank 5 of 21
What it means and the source
How much text the model keeps in view in a single request, question and answer included. It matters for long documents and whole code projects. A token is a fragment of a word.
huggingface.coOpens the source in a new tab Oct 23, 2026
Knowledge and abilities
- Reads images
- Yes13 models out of 23
What it means and the source
The model can take images (photos, screenshots, charts) as well as text.
huggingface.coOpens the source in a new tab Oct 23, 2026
Size and openness
- Parameters
- 2.8 trillion
What it means and the source
The size of the network: how many billions of learned values make it up. Bigger does not necessarily mean better, but it tells how much hardware it takes to run it yourself.
huggingface.coOpens the source in a new tab Oct 23, 2026- Active parameters per token
- 104 billion
What it means and the source
In «mixture of experts» models only part of the network works on each word: these are the parameters actually used for each token, and they matter for speed and running costs.
huggingface.coOpens the source in a new tab Oct 23, 2026- Open weights (downloadable)
- Yes7 models out of 23
What it means and the source
The model's weights can be downloaded: you can run it on your own servers, adapt it and decide where the data stays, within the licence terms.
huggingface.coOpens the source in a new tab Oct 23, 2026
Where it stands · Kimi K3
The same verified figures, next to those of the other catalogue models. Tap a model for the head-to-head.
Hardware to try AI at home
Kimi K3: just holding the weights in memory takes about 5,600 GB at 16-bit, 2,800 GB at 8-bit, 1,400 GB at 4-bit. Beyond 24 GB it is not a home-PC job: using it through an API makes more sense.Worked out from the billions of parameters the vendor states, not a measurement: context and the runtime need more memory.
Questions and answers
Kimi K3: how much text can it read at once?
The context window is 1,048,576 tokens.
Kimi K3: can I run it on my own computer?
Yes, the vendor publishes the model's weights. Parameters: 2.8 trillion. Holding them in memory takes about 1,400 GB at 4-bit; context and the runtime need more memory.
L’essentiel
Sur Moonshot AI Kimi K3, la donnée qui compte vraiment est Fenêtre de contexte : 1048576. Je l’ai relevé chez Moonshot AI – model card (Hugging Face). Parmi les 21 que j’ai vérifiés, 4 font mieux. L’autre chiffre à regarder est Lit les images : 1.
Les chiffres de Moonshot AI Kimi K3
Je les ai classés par importance, avec la valeur et la position face aux autres que j’ai vérifiés.
Fenêtre de contexte : 1048576. Position 5 sur les 21 que j’ai comparés.
Lit les images : 1. C’est la meilleure valeur parmi les 13 que j’ai comparés.
Open weights (downloadable) : 1. C’est la meilleure valeur parmi les 7 que j’ai comparés.
Paramètres (milliards) : 2.8T. C’est la meilleure valeur parmi les 7 que j’ai comparés.
Paramètres actifs par token (milliards) : 104B. C’est la meilleure valeur parmi les 2 que j’ai comparés.
À qui il s’adresse
Choisis-le si Lit les images est ta priorité : 1, et aucun de ceux que j’ai vérifiés ne fait mieux.
Il te convient si tu cherches Open weights (downloadable) : 1.
Regarde ailleurs si ce qui compte pour toi est Fenêtre de contexte : ici c’est 1048576, et 4 produits font mieux.
Pour aller plus loin
Sources
Ask Sir Compare:Explain these numbers to me
Reviews and tests
Moonshot AI Kimi K3
Reviews and tests · Moonshot AI Kimi K3
Reviews are not available yet.
What changed
- Page updated
Sir Compare’s letter
Updates on AI Models & LLMs? I’ll send them your way.
Pick what you care about: I only write when something new comes out — a guide, a news story, a verified price that drops.
Step 1 of 2
What are you into?
Pick as many as you like. Pick none and you get what’s new in every section.
- Only what you pick
- Never an empty email
- One click to leave