deep-principle/Hydrogen-1.0-MXFP4

MXFP4 quantized version of Hydrogen-1.0

Tech blog · Usage · Versions · 中文说明

Overview

Hydrogen-1.0 is a scientific problem-solving and research assistance model developed by Deep Principle based on GLM-5.3. Through text interaction, it is intended to assist with scientific questions, explain derivations, and organize responses to open-ended research problems. Outputs should be assessed against the assumptions, constraints, and evidence relevant to each task.

Research in practice

The technical blog presents examples of scientific computing, reaction-network exploration, NMR structure analysis, and experimental planning. These demonstrations combine the model with external tools and specific research environments; outputs require scientific review.

Versions

Model details

This repository is dedicated to the MXFP4 quantized version of Hydrogen-1.0. Available context and output budgets depend on the inference framework, service configuration, and runtime resources. Use a runtime that supports the exported MXFP4 checkpoint format and follow the deployment requirements for these weight files.

Quantization and deployment

Quantized with AMD Quark. Dense MLPs, shared and routed experts, and MTP MLPs use MXFP4 with dynamic activations; attention, router, embeddings, lm_head and eh_proj remain BF16. The repository's quantization configuration uses groups of 32 elements with E8M0 scales for weights and activations.

The team reports testing on AMD Instinct MI350X with ATOM at tensor parallel 4, FP8 KV cache, and speculative decoding off. The following example assumes a compatible ATOM installation and does not enable a speculative decoding method:

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

ATOM uses --server-port for HTTP; --port is internal engine communication. The API model ID is hydrogen-1.0. Available context and output budgets depend on runtime, service settings and memory; this command does not establish 1M-context operation on every setup.

Monitor KV cache utilization and eviction under load. In the team's reported TP4 test, 128 concurrent long reasoning requests filled the pool and triggered eviction, with a sharp drop in answer quality; at 48 concurrent requests the pool stayed below 75% and quality was normal. These are deployment-specific observations, not universal concurrency limits.

Keep the matching end-of-turn settings in generation_config.json, along with the tokenizer, chat template, model and quantization configurations, and weight index. The reported deployment relies on these settings for correct stopping.

Usage example

The following example calls an already running compatible inference service. The launch command above sets the API model ID to hydrogen-1.0; confirm it at /v1/models and set the base URL for your deployment. This example does not download or load weights, and does not imply that a hosted endpoint is available on either model hub.

python -m pip install openai
export MODEL_BASE_URL="http://localhost:8001/v1"
export SERVED_MODEL_NAME="hydrogen-1.0"
# Set MODEL_API_KEY separately if the service requires authentication.

import os
from openai import OpenAI

client = OpenAI(
    base_url=os.environ["MODEL_BASE_URL"],
    api_key=os.environ.get("MODEL_API_KEY", "not-required"),
)

stream = client.chat.completions.create(
    model=os.environ["SERVED_MODEL_NAME"],
    messages=[
        {"role": "system", "content": "You are an AI scientist. State assumptions and explain the scientific reasoning behind your answer."},
        {"role": "user", "content": "Derive the period of a simple pendulum under the small-angle approximation, and explain when the approximation breaks down."},
    ],
    max_tokens=4096,
    stream=True,
    extra_body={"chat_template_kwargs": {"reasoning_effort": "max"}},
)

finish_reason = None
has_answer = False
for chunk in stream:
    if not chunk.choices:
        continue
    choice = chunk.choices[0]
    if choice.finish_reason is not None:
        finish_reason = choice.finish_reason
    text = choice.delta.content or ""
    if text:
        has_answer = has_answer or bool(text.strip())
        print(text, end="", flush=True)
print()
if finish_reason == "length":
    raise RuntimeError('The response was truncated. Increase max_tokens within the service limits.')
if not has_answer:
    raise RuntimeError('No final answer was returned. Check the request status and output budget.')

The example uses an illustrative per-request budget of max_tokens=4096, rather than the model's maximum output length. Adjust it for complex reasoning tasks within the service limits. The code displays the final-answer content from the stream. See usage notes for additional request conventions.

Intended uses

• Research and educational exploration of scientific questions.
• Assistance with explaining equations, derivations, assumptions, and possible solution approaches.
• Drafting structured responses to open-ended scientific problems for expert review.
• Studying model behavior and developing downstream research tools within the applicable license terms.

Model outputs are supporting material for users to review. They should not be treated as verified experimental findings or as sufficient grounds for decisions affecting health, safety, or other consequential outcomes.

Limitations and biases

The model may produce incorrect facts, calculations, equations, references, or scientific explanations. A detailed or confident response does not establish correctness. Users should verify key claims with primary literature, independent calculations, or experiments as appropriate.

Performance may vary across disciplines, languages, task formats, and input lengths. Training data may introduce selection, cultural, or linguistic biases. Outputs should not be treated as a complete or unbiased representation of scientific consensus.

Long-context capacity does not guarantee reliable retrieval, understanding, or use of every input detail. MXFP4 quantization may affect outputs. Validate the quantized model for the intended task; this statement does not imply a measured accuracy change for this release.

Training data and procedure

The technical blog describes the public training overview, including supervised fine-tuning (SFT), reinforcement learning (RL), and data-source categories such as domain knowledge, research scenarios, laboratory records, and discovery trajectories. Full training datasets and complete training recipes are not released with the model weights in this repository.

Evaluation

Evaluation focuses on scientific multiple-choice questions and open-ended scientific research problems. Validate the model for the intended task, inputs, and inference configuration. Model answers do not replace independent verification of scientific conclusions.

License and attribution

Hydrogen-1.0 is developed by Deep Principle based on GLM-5.3 by Z.AI. The Hydrogen-1.0 MXFP4 quantized model weights in this repository are provided under the MIT License. Preserve the included copyright and permission notices when using or redistributing the weights.

Independently authored Deep Principle code and documentation are licensed under Apache License 2.0. Third-party files retain their respective licenses. See licensing scope and attribution notices for details.

Contact

Deep Principle: info@deepprinciple.com