Skip to content

Reflection Beam open weights: a 501B MoE with 1M context on Hugging Face

Reflection Beam open weights are out: 501B parameters, 23B active, Apache 2.0 license and a 1M-token context. What it can do and how to try it.

Piotr Sobolewski3 min readUpdated

This article in other languages:العربيةDeutschEspañolFrançaisItaliano日本語Polski简体中文

In this article
  1. What exactly is in the Reflection Beam open-weights release?
  2. What can Reflection Beam do?
  3. Why all the buzz around Beam?
  4. How to try Reflection Beam?
  5. FAQ

On October 16, Reflection AI released the Reflection Beam open weights on Hugging Face under the Apache 2.0 license. Beam is a MoE (mixture-of-experts) model with 501B parameters, 23B of which are active for each token. The company announced it on October 5.

What exactly is in the Reflection Beam open-weights release?

It’s not just the weights. The package includes:

  1. Weights in three variants: BF16, FP8 and NVFP4, plus a small drafter model for speculative decoding (it speeds up token generation).
  2. Fused kernels and ready-made serving configs.
  3. Day-one support in Transformers, vLLM, SGLang, Unsloth, Axolotl and NVIDIA Dynamo.
  4. An open-source agent harness, Mirror CLI (Python frontend, Rust core), with parallel agents in separate worktrees, MCP and skills.
  5. Fine-tuning and RL recipes, with self-distillation recipes coming “in the coming days”, according to the company.

On top of that there’s a model card and a safety report. Reflection says the launch is backed by more than 40 partners, from AWS, Google Cloud and Azure to OpenRouter, Ollama and LM Studio.

The biggest deal, though, is the license. Apache 2.0 means full freedom, commercial use included, without the extra conditions that come with, say, Meta’s Muse Spark Open license.

What can Reflection Beam do?

It’s a text model (no images, no audio) built for coding, reasoning and long agentic tasks. The context reaches 1M tokens, and its knowledge cutoff is June 30, 2026.

Results from the model card for the BF16 version:

Benchmark Score
SWE-bench Verified 81.6%
Terminal Bench v2.1 80.1%
LiveCodeBench 88.3%
GPQA Diamond 90.6%
AIME 2026 97.7%
HLE (with tools) 51.1%

These are Reflection’s own numbers and nobody has verified them independently yet, so let’s take them with a grain of salt. Interestingly, the FP8 and NVFP4 versions score almost the same as BF16 on these tests.

There’s one small caveat: according to the model card, Beam was built “predominantly for English”.

Why all the buzz around Beam?

First, the scale of training. According to the announcement post, Beam was pretrained on 23.8 trillion tokens on 6,144 NVIDIA GB300 GPUs in under four weeks. RL took another four weeks on 10,500 GB300s, with more than 100 million rollouts and about 1.3 billion sandboxes. At peak, the company used 883,006 vCPUs across 23 clusters on GCP and AWS.

Second, who’s behind it. According to TechCrunch, Reflection is a Brooklyn startup founded in 2024 by two former Google DeepMind researchers. It has raised about $4.7 billion (PitchBook data), from backers including Nvidia and Sequoia.

Third, the direction. According to the announcement, Beam is meant to push the “Western open-weight frontier” forward, and TechCrunch frames the company as the Western answer to Chinese open models from DeepSeek, Qwen and Z.ai. Beam is the first model in the series, so this looks more like a beginning than a one-off.

How to try Reflection Beam?

The easiest way is the API. The Reflection platform, in public beta, is “free to start” according to the company. The model is also on OpenRouter as reflection/beam-1.0-501b-a23b.

For coding, Beam was tested with OpenCode, Pi and Hermes, among others, and Reflection has a guide to hooking it up to agent harnesses.

What about running it locally? The BF16 weights alone are about 1 TB (501B parameters × 2 bytes), so it’s more of a toy for a server room than for a home PC ;) Ollama and LM Studio are on the partner list, but your hardware still has to handle that size.

FAQ

Is Reflection Beam free?

The weights are: you can download them from Hugging Face and use them under the Apache 2.0 license, commercially too. The Reflection API in public beta is “free to start” according to the company, but no limits have been given.

Does Beam work in languages other than English?

According to the model card, Beam was developed mainly for English, and no list of supported languages is given. It’s worth checking on your own prompts.

Does Beam handle images?

No. It’s a text-only model. It can handle other formats only if we convert them to text first.