GigaChat3.5-432B-A28B

GigaChat3.5-432B-A28B

Generates text in Russian and English for chat, code, reasoning and tool use. A 432B Mixture-of-Experts model with 28B active parameters, aimed at teams that host large models.

Made by ai-sage on Hugging Face. These notes are Mindket's, written from its model page and model card; the model's own files and license are on Hugging Face.

About

What it is: GigaChat 3.5 Ultra is a large language model from ai-sage that writes text for assistant chat, reasoning, code and tool-use work. It is a Mixture-of-Experts model with 432B parameters in total, and 28B of them are active for each token. It handles Russian and English. The card says it is about 40% smaller than the 700B GigaChat 3.1 Ultra and ahead of it in code, math and agent tasks.

Who it suits: Teams that host large open-weight models and need Russian and English answers fit best. Developers who build coding assistants or agents that call tools can use it too, and so can researchers who study attention designs.

Industry | Who uses it | How they use it | Example
Software | Developers | Run it behind a coding assistant or an agent that calls tools | Asking it to explain a function and suggest a fix
Research | ML researchers | Study the hybrid attention design and compare it with other large models | Reading the architecture sections and benchmark tables in the model card
Hosting | Infrastructure teams | Serve the fp8 weights on a large cluster | Trying the two prediction heads to see how fast replies come
Language services | Teams working in Russian and English | Build assistants that answer in either language | A support bot that replies in Russian or in English

You can skip it if you only want a small model for a laptop, because the weights alone take about 434 GB.

Where to use it: The model runs wherever you can fetch the files and serve a model of this size.

• Windows, macOS or Linux: download the files with Python and the Hugging Face tools.
• Large clusters: the card targets cluster deployment with fp8 inference.
• Assistant and agent apps: the card names multilingual assistants and tool-use scenarios.
• Code workflows: code is one of the named target workloads.

Use case | Who | What they get back
Multilingual assistant | Product teams | Replies in Russian or English
Code help | Developers | Code and explanations written by the model
Agent steps | Agent builders | The model's text output inside a tool-using loop
Research comparison | ML researchers | A large open model to set beside others

How to set it up: You need three things before you start.

• Python on Windows, macOS or Linux
• About 434 GB of free disk for the fp8 weights, or about 217 GB at 4-bit, which is an estimate
• Hardware for a 432B model, since the card aims at large-cluster deployment

1. Install Python if you do not have it.
2. Install the Hugging Face hub library.
3. Download the model files with the hf command.
4. Follow the model card to run the weights, since that is where the run steps live.

pip install -U huggingface_hub
hf download ai-sage/GigaChat3.5-432B-A28B

How to use it: Once the files are on disk, load them in your own inference setup as the model card describes, then send prompts in Russian or English. The command below fetches the fp8 model files, and the table lists the other copies the card points to.

hf download ai-sage/GigaChat3.5-432B-A28B

Tool | What it does
GigaChat3.5-432B-A28B-bf16 | The same model in bf16 precision
GigaChat3.5-432B-A28B-GGUF | A GGUF version of the model
GigaChat3.5-432B-A28B-base | The base version meant for training
GigaChat3.5-432B-A28B-checkpoints | Training checkpoints

Example: You ask: "Write a Python function that removes duplicate lines from a text file." The model sends back the function as text, with a short note on how it works.

For newcomers and experts: A language model like this reads your text and writes the reply, which makes it the part of a chat assistant that does the answering. It is huge. The first small thing to try is opening the model page and comparing the listed file sizes with your free disk space.

Only 28B of the 432B parameters work on each token, so serving costs far less than a dense model of the same size. Some layers use Multi-head Latent Attention and the rest use GatedDeltaNet linear attention, which lowers the cost of long context. The card reports about 4x less KV-cache per token than 3.1 Ultra and about 20% higher generation throughput. Two Multi-Token Prediction heads speed up greedy decoding by up to 2.2x. Training ran in native FP8, and alignment went through SFT, DPO and online RL. The trade-off is size, since the weights alone take about 434 GB.

Download the fp8 weights
Load them on your own hardware
Hybrid attention layers read your prompt
28B of 432B parameters work per token
Two MTP heads speed up decoding
The reply text comes back

Good to know: The scores in the card cover the base model, not this chat version, and they compare average results on general tasks.

• Precision: this repository holds fp8 weights, and a bf16 copy sits in a separate repository.
• Languages: the record lists Russian and English.
• Other versions: the record counts one quantized version made by other people.

Title: Average base-model score on general tasks
GigaChat 3.1 Base (700B): 69.6 points
GigaChat 3.5 Ultra Base (430B): 72.3 points
DeepSeek V4 Flash Base (284B): 71.9 points
DeepSeek V3.2 Exp Base (685B): 71.5 points

• License: MIT.
• Cost to run: no API key, account or paid service is named, but the weights need about 434 GB of disk and large hardware.
• Data it reaches: it downloads files from Hugging Face, and the model runs on your own machines.
• Not for: the source names no forbidden uses, and the full weights need about 434 GB.

On Hugging Face, 13 November 2026

• Task: text generation
• Parameters: 434B
• Weights: about 434 GB in F8_E4M3; about 217 GB at 4-bit (estimate)
• Formats: Safetensors
• Downloads in the last 30 days: 654
• Likes: 84
• License: mit
• Languages: ru, en
• Based on: ai-sage/GigaChat3.5-432B-A28B-base
• Quantized versions: 1
• Tags: safetensors, gigachat3_5, instruct, multilingual, long-context, tool-use, hybrid-attention, linear-attention, text-generation, conversational, fp8

Links

• Model: https://huggingface.co/ai-sage/GigaChat3.5-432B-A28B
• Files: https://huggingface.co/ai-sage/GigaChat3.5-432B-A28B/tree/main

Get the model

Download its files with Hugging Face's command line:

pip install -U huggingface_hub
hf download ai-sage/GigaChat3.5-432B-A28B

Then follow the usage in its model card. Nothing in this kit runs by itself.