Hydrogen-1.0-MXFP4

Hydrogen-1.0-MXFP4

An MXFP4 quantized copy of Deep Principle's 384B Hydrogen-1.0 science assistant, for AMD Instinct hardware. For labs that want a text model for scientific questions and research planning.

Made by deep-principle 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: Hydrogen-1.0-MXFP4 is the MXFP4 quantized release of Hydrogen-1.0, a text model from Deep Principle built on GLM-5.3. It is meant to help with scientific questions, explain derivations and organize answers to open research problems. The record lists 384B parameters and an MIT license.

Who it suits: Research teams with large AMD Instinct hardware who want an assistant for scientific reasoning.

Industry | Who uses it | How they use it | Example
Chemistry | Research chemists | Explore reaction networks and plan experiments | A question about a reaction pathway
Materials science | Lab teams | Work through derivations and calculations | A step-by-step derivation
Academia | Researchers | Draft analyses of open research problems | A structured plan for a study
AI platforms | Infrastructure teams | Serve a large open model on AMD hardware | An internal science assistant

You can skip it if you have no AMD Instinct hardware, since this build targets it. You can also skip it for non-scientific work.

Where to use it:

• ATOM on AMD Instinct MI350X, which the team reports testing
• An inference service you run yourself, called through an API
• Windows, macOS or Linux, with Python for the download step

Use case | Who | What they get back
Science questions | Researchers | Explanations and worked reasoning, to be reviewed
Experiment planning | Lab teams | Structured plans for experiments
Long documents | Analysts | A context of 1M tokens, depending on your setup

How to set it up:

• Python and pip on your computer
• AMD Instinct hardware with an ATOM install, as the team tested
• A runtime that supports the MXFP4 checkpoint format
• Roughly 384 GB of weight files

1. Download the files with the hf command.
2. Install a compatible ATOM runtime.
3. Launch the OpenAI-style server with the card's command.
4. Confirm the model ID at the models endpoint of your service.

pip install -U huggingface_hub
hf download deep-principle/Hydrogen-1.0-MXFP4

How to use it: Start the server, then call it like any chat API with the model ID hydrogen-1.0. The card says the server port flag sets HTTP and the plain port flag is for engine communication.

python -m atom.entrypoints.openai_server \
  --model deep-principle/Hydrogen-1.0-MXFP4 \
  --served-model-name hydrogen-1.0 \
  --trust-remote-code --kv_cache_dtype fp8 -tp 4 \
  --host 0.0.0.0 --server-port 8001

Example: You ask a question about a reaction mechanism through the chat API. The model replies with an explanation and reasoning. The card stresses that outputs need scientific review against the task's assumptions.

For newcomers and experts: If you are new, this model needs serious hardware, so read the card's deployment notes before you try it. Treat its answers as drafts that a specialist checks.

If you know the field, the card says attention, router, embeddings and lm_head stay BF16 while MLPs and experts use MXFP4 with groups of 32 and E8M0 scales, quantized with AMD Quark. In its TP4 test, 128 concurrent long requests triggered KV cache eviction and quality fell, while 48 stayed normal.

Hydrogen-1.0 base model
AMD Quark MXFP4 quantization
ATOM server on AMD Instinct
Chat API answers

Title: Concurrent long requests in the team's TP4 test
Normal quality: 48
Cache pool filled and quality fell: 128

Good to know:

• License: MIT
• Cost to run: your own AMD Instinct hardware, with no paid service named
• Data it reaches: the files you download and the prompts you send your own server
• Not for: unreviewed scientific claims, since the card says outputs need scientific review
• Limits: the 1M context and 128K output depend on your runtime and memory

On Hugging Face, 13 November 2026

• Task: text generation
• Parameters: 384B
• Weights: about 384 GB in U8
• Formats: Safetensors
• Downloads in the last 30 days: 12,804
• Likes: 16
• License: mit
• Based on: deep-principle/Hydrogen-1.0
• Tags: safetensors, glm_moe_dsa, mxfp4, science, text-generation, conversational, 8-bit, quark

Links

• Model: https://huggingface.co/deep-principle/Hydrogen-1.0-MXFP4
• Files: https://huggingface.co/deep-principle/Hydrogen-1.0-MXFP4/tree/main

Get the model

Download its files with Hugging Face's command line:

pip install -U huggingface_hub
hf download deep-principle/Hydrogen-1.0-MXFP4

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