Skip to content
Computer vision

HiTZ/Latxa-3-122B-A10B

a multimodal language model for Basque

2 min readupdated

Latxa 3 is a large multimodal instruct model (trained to follow instructions) created by researchers at the HiTZ center and the IXA group (University of the Basque Country). Built on Qwen3.5-122B-A10B using a Mixture-of-Experts (MoE) architecture, it has 122 billion total parameters with roughly 10 billion active per token and supports context windows up to 262,144 tokens. Latxa 3 handles images and holds conversations in Basque, operating exclusively in thinking mode to reason through responses before answering. It was built to close the AI performance gap for low-resource languages.

Key features

  • Support for image inputs and multimodal instructions.
  • Mandatory thinking mode that reasons through answers before responding.
  • Significant accuracy gains on Basque language and cultural knowledge benchmarks over the base model.
  • Massive context window of up to 262,144 tokens.

What you can do

  • Deploy a Russian or Basque AI assistant for enterprise or research workflows.
  • Run multimodal image analysis with text prompts on top of a capable open-weights architecture.
  • Integrate the model into services via an OpenAI-compatible API using frameworks like vLLM.

Example

Sample request passing an image to the model through an OpenAI-compatible client:

from openai import OpenAI

client = OpenAI(base_url="http://localhost:8000/v1", api_key="EMPTY")

messages = [
    {
        "role": "user",
        "content": [
            {"type": "image_url", "image_url": {"url": "[https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/p-blog/candy.JPG](https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/p-blog/candy.JPG)"}},
            {"type": "text", "text": "Zer ikusten da irudi honetan?"},
        ],
    },
]

response = client.chat.completions.create(
    model="Latxa-3",
    messages=messages,
    max_tokens=1024,
)
print(response.choices[0].message.content)
Read moreDetails
Built on the Qwen3.5-122B-A10B architecture (MoE: 256 experts, 8 routed + 1 shared per token).
Scored an average of 84.0% on Basque language proficiency tasks and 76.4% on Basque knowledge benchmarks.
Hardware requirements: BF16 weights require roughly 245 GB of VRAM (e.g., 4x 80 GB GPUs). An FP8 version is also available.
License: Apache 2.0.
Discuss Telegram 2 34

Will it run?

Won't run

Not enough memory even for 4-bit. Below are GPUs you can rent.

  • BF16weights 228 GB · ready files252 GB
  • FP8 / INT8recommendedweights 114 GB126 GB
  • 4-bitweights 64 GB71 GB

Tip: go 8-bit.FP8 / INT8 is within about a percent of full precision in quality, takes half the memory and runs faster on RTX 40/50 and H100. Full precision only pays off for fine-tuning.

What to get for this model

Massed ComputeOpen Massed Compute

Your own GPU: Won't fit any home GPU — renting is cheaper.

Massed Compute hourly prices, November 2026. The discounted price applies with the coupon at checkout.

Estimate: weights plus working memory. Long videos, high resolutions and long contexts need more.

Run it on Massed Compute

The script downloads the model to the rented machine. How to run it is in the model card.

  1. 1. Rent a machine with a fitting GPU

    Open Massed Compute
  2. 2. Paste the script into the machine's terminal

    Connect over SSH (the Massed Compute panel shows the command) or open a terminal on the machine's desktop.

    HF_TOKEN / CIVITAI_TOKENThe script starts with token lines: paste yours between the quotes. The Hugging Face token is for gated models, the Civitai one for Civitai models.

    # HiTZ/Latxa-3-122B-A10B
    mkdir -p ~/models
    curl -LsSf https://astral.sh/uv/install.sh | sh
    source "$HOME/.local/bin/env"
    uv tool install "huggingface_hub[cli]"
    hf download HiTZ/Latxa-3-122B-A10B --local-dir ~/models/Latxa-3-122B-A10B
    # The model is in ~/models. How to run it is in the model card (link above).
    Add more LoRAs to Latxa-3-122B-A10B
  3. 3. Use it

    The model is in ~/models. How to run it is in the model card (link above).

Stop the machine in the panel when you are done: it bills by the minute while it runs.

How to speed it up

Picked for this model type. Choose your hardware above to see what fits you exactly.

  • FlashAttention 44.0.0b40from November 18, 2026new release

    fastest exact attention on B200

    Kernels for Blackwell (beta). Exact attention without quantisation — fine for everything, LLMs included.

    How to enable: pip install flash-attn-4; vLLM and SGLang pick it up themselves.

    Earlier generations (2)
    • FlashAttention 3 — The Hopper (H100/H200) version with FP8.
    • FlashAttention 2 — The standard for RTX 30/40/50 and A100: exact attention without the memory cost of long contexts.
  • torch.compile2.15.0from October 28, 2026new release

    10–30% for free

    Compiles the model for your GPU. Slower first run, faster afterwards.

    How to enable: model = torch.compile(model); in ComfyUI the TorchCompileModel node.

Model family

New posts about versions of Qwen3.5-122B-A10B (quants, finetunes, LoRAs) will show up under the bell.

Quantised versions 12All on Hugging Face
Fine-tunes 12All on Hugging Face
LoRAs and adapters 2All on Hugging Face

About the model

HiTZ/Latxa-3-122B-A10B
Author
HiTZ
Category
Computer vision
Released
Posted

On NeyroLearn

Hugging Face

342

Ask AI about this model

Answered by Gemini 4 · the chat is kept only in this tab

Comments