Beam-1.0-501B-A23B-NVFP4

Beam-1.0-501B-A23B-NVFP4

Reads and writes text with a 1M-token context, built for coding, agent tool use and math reasoning, as a 501B-parameter mixture-of-experts model in NVFP4 form. Suits teams with multi-GPU servers who want an open model.

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-NVFP4 is the NVFP4-quantized, instruction-tuned checkpoint of Beam, an open model from Reflection. Beam takes text in and gives text back. It is a sparse mixture-of-experts model with 501 billion parameters in total and 23 billion active per token, built mainly for English.

Who it suits: Engineering and research teams with large GPU servers who want an open model for coding, agent work and reasoning, under the Apache 2.0 license.

Industry | Who uses it | How they use it | Example
Software | Engineering teams | Fix real GitHub issues and work in command-line and DevOps setups | Resolve a bug report in a repository
AI products | Agent builders | Run long multi-step tasks with MCP tool use and browsing | Carry out a workflow from a plain request
Science and education | Researchers | Work through graduate-level science, proofs and applied maths | Check a proof step by step
Infrastructure | Platform teams | Serve one large open model on their own GPUs | Host an internal coding assistant

Skip it if you have a single GPU or a laptop, since the weights alone come to about 255 GB.

Where to use it: The card points to agent-style coding tools and to its own documentation.

• Agentic coding tools and command-line workflows
• Servers that can run a huge mixture-of-experts model
• Windows, macOS or Linux, with Python for the download

Use case | Who | What they get back
Software engineering | Developers | Code changes for real issues, in many languages
Agentic tool use | Agent builders | Finished multi-step tasks that use MCP tools and browsing
Math and science | Researchers | Worked solutions and proofs

How to set it up: You need Python, a lot of disk space and a multi-GPU server.

• Python with pip
• About 255 GB of disk for the weights
• A serving setup that supports this model, as the Reflection documentation describes

1. Install the Hugging Face hub tools with pip.
2. Download the model files with the hf command.
3. Follow the Reflection documentation to serve the checkpoint.

pip install -U huggingface_hub
hf download reflection/Beam-1.0-501B-A23B-NVFP4

How to use it: Serve the checkpoint, then send chat requests that use its chat template, a modified version of the Harmony format. The card text shown here gives no code, so follow the Reflection documentation for the serving steps.

Example: You ask the served model to read a bug report from a repository and propose a fix. It replies with reasoning and a code change in text, since text is its only input and output.

For newcomers and experts: If you are new to large models, know that this one needs a server-class machine. Start by reading the Reflection documentation, and treat the download as the last step.

If you know the field, the model has 52 layers, 3 dense and 49 mixture-of-experts, with 256 experts, gated grouped-query attention and sliding window attention. It has a 152k-token vocabulary and a 1M-token context. The card claims comparable solution quality with about 3x fewer reasoning tokens than leading open models.

Download files with hf
Serve the NVFP4 checkpoint
Send chat-template messages
Route tokens through active experts
Receive reasoning and a text reply

Good to know: The release also includes BF16 and FP8 checkpoints, and this one is the NVFP4 version. The card lists a technical report as coming.

• License: Apache 2.0
• Cost to run: none named, apart from your own hardware
• Data it reaches: the model files on your computer, plus any tools or browsing you connect to an agent setup
• Not for: images or audio, since the supported modality is text only

On Hugging Face, 13 November 2026

• Task: text generation
• Parameters: 255B
• Weights: about 255 GB in U8
• Formats: Safetensors
• Downloads in the last 30 days: 6,615
• Likes: 45
• License: apache-2.0
• Languages: en
• Tags: safetensors, beam, mixture-of-experts, reasoning, code, tool-use, nvfp4, text-generation, conversational, 8-bit, modelopt

Links

• Model: https://huggingface.co/reflection/Beam-1.0-501B-A23B-NVFP4
• Files: https://huggingface.co/reflection/Beam-1.0-501B-A23B-NVFP4/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-NVFP4

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