Beam-1.0-501B-A23B

Beam-1.0-501B-A23B

An open instruction-tuned text model from Reflection with 501B parameters, 23B active at a time and a 1M-token context, aimed at coding, tool use and reasoning. Needs a large multi-GPU server to run.

Made by reflection 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: Beam-1.0-501B-A23B is an open model from Reflection that takes text in and writes text out. It uses a sparse mixture-of-experts design with 501B total parameters and 23B active for each token. The card describes a context window of 1M tokens and says it was developed mostly for English. This release is the instruction-tuned model, offered in BF16, FP8 and NVFP4 forms.

Who it suits: Teams with large GPU servers who want a big open model for coding agents, tool-using agents and hard reasoning, and researchers who study large open models.

Industry | Who uses it | How they use it | Example
Software engineering | Engineering teams | Fix issues and work in terminal and DevOps setups | A coding agent that opens a pull request
Agent building | Agent developers | Run long, multi-step tasks with tools | A browsing agent that finishes a research task
Research | ML researchers | Study a large sparse model and its reasoning | Testing how it spends reasoning tokens
Science and maths | Analysts | Work through graduate-level problems | A step-by-step proof attempt

If you only have a laptop or one consumer GPU, you can skip it, because the weights alone take about 1 TB in BF16.

Where to use it:

• Large GPU servers that can hold the weights
• Coding agent setups, where the card says it connects directly
• Tool and browsing workflows, including MCP-based tools

Use case | Who | What they get back
Software tasks | Engineers | Code, fixes and command-line steps
Agent work | Agent developers | Multi-step results from plain-language instructions
Reasoning | Researchers | Worked solutions to maths and science problems

How to set it up:

• Python and pip
• About 1,003 GB of disk space in BF16, or about 251 GB at 4-bit (an estimate)
• A multi-GPU server sized for a model this large, which the card's visible text does not detail

1. Install the Hugging Face hub tools.
2. Download the model files in the format your servers support.
3. Follow the card and Reflection's documentation to serve it.

pip install -U huggingface_hub
hf download reflection/Beam-1.0-501B-A23B

How to use it: The part of the card shown here has no run command, so follow its serving notes and the maker's documentation. The model takes a chat-style text prompt and replies with text. The card says a technical report is still to come.

Example: You point a coding agent at the model and give it a bug report from a real project. According to the card, the model can read the issue, work in a command-line setup and propose a fix, though the card shows no sample run here.

For newcomers and experts: If you are new to this, try a smaller open model first, since this one needs a huge server. Experts can compare its reasoning cost with other open models, which the card says is its main selling point.

The card lists 52 layers, 3 dense and 49 mixture-of-experts, with 256 experts, a 152k-token vocabulary and gated grouped-query attention mixed with sliding-window attention. It says Beam reaches answer quality similar to leading open models while using about three times fewer reasoning tokens. Treat that as the maker's claim until you test it.

You send a text prompt or task
The router picks a few experts per token
The model reasons, using tools if set up
It writes the answer or the next step

Good to know:

• Language: developed mostly for English
• Input: text only, with no images or audio
• Variants: the Hugging Face record lists 12 quantized and 5 fine-tuned versions made by others

Title: Disk size of the weights, in GB
BF16: 1,003 GB
4-bit estimate: 251 GB

• License: apache-2.0
• Cost to run: no API key or paid service is named for self-hosting, but the model needs a huge GPU server
• Data it reaches: downloading fetches model files, and the model reads the text you give it. With browsing or tool setups turned on, it can reach whatever those tools reach
• Not for: laptops, single small GPUs, image or audio input, or work that must be mainly in another language

On Hugging Face, 13 November 2026

• Task: text generation
• Parameters: 501B
• Weights: about 1,003 GB in BF16; about 251 GB at 4-bit (estimate)
• Formats: Safetensors
• Downloads in the last 30 days: 13,747
• Likes: 319
• License: apache-2.0
• Languages: en
• Quantized versions: 12
• Fine-tunes: 5
• Tags: safetensors, beam, mixture-of-experts, reasoning, code, tool-use, text-generation, conversational

Links

• Model: https://huggingface.co/reflection/Beam-1.0-501B-A23B
• Files: https://huggingface.co/reflection/Beam-1.0-501B-A23B/tree/main

Get the model

Download its files with Hugging Face's command line:

pip install -U huggingface_hub
hf download reflection/Beam-1.0-501B-A23B

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