Skip to content
AI Radar by Wagner Solution
CS
Menu
major New models verified update

Reflection AI releases Beam with open weights under the Apache 2.0 license

confirmed by 2 independent sources updated 14 h ago

Beam is now available under the Apache 2.0 license with support from the transformers, SGLang and vLLM libraries. It offers a million-token context and five reasoning levels for coding and agentic tasks. It has 501 billion parameters, of which it activates 23 billion per token.

Reflection AI has released Beam with open weights under the Apache 2.0 license. The article on Hugging Face Blog already describes the available checkpoints in BF16, FP8 and NVFP4 formats and support from the transformers, SGLang and vLLM libraries. The model also comes with an open harness, an environment for running tasks, written in Rust. Availability has thus moved from the previously announced release during the month and access for selected users to publicly released weights.

Beam processes only text and focuses on coding, reasoning and agentic tasks. Its mixture-of-experts architecture has 501 billion parameters in total but activates 23 billion for each token. The context window holds one million tokens. The model supports tool calls, preserving the reasoning process between conversation turns, and five reasoning levels: low, medium, high, xhigh and max; the default is medium.

The lower number of active parameters does not eliminate the memory requirements of the full model: all experts remain in memory during deployment. According to Hugging Face Blog, the BF16 checkpoint is approximately 1 TB. The NVFP4 variant stores most parameters at four-bit precision and is designed for NVIDIA Blackwell GPUs; it retains some parameters in BF16 for accuracy.

According to Reflection AI, Beam achieves results comparable to GLM-5.2 on demanding reasoning tasks with inference compute requirements three to four times lower. The TechCrunch article notes that these claims have not been independently verified. The company reports pretraining on 23.8 trillion tokens, followed by reinforcement learning lasting more than four weeks on 10 500 Nvidia GB300 GPUs.

What changed

Why it matters

The released weights allow teams to run and modify the model for their own coding and agentic tasks on their own infrastructure or in the cloud. When deciding on deployment, however, it is necessary to distinguish compute requirements from memory capacity: 23 billion active parameters does not mean that hardware sufficient for a standalone model of that size will suffice. The claimed compute savings need to be verified on the specific workload.

Release card

Beam

Reflection AI

open weights
Specifications
The model has 501 billion parameters in total and activates 23 billion for each token.
Context
The context window holds 1 million tokens.
Inputs
The model accepts text and generates text.
Licence
Apache 2.0
Availability
The weights are publicly available on Hugging Face in BF16, FP8 and NVFP4 formats. The model is supported by the transformers, SGLang and vLLM libraries. The release includes an open environment for running agentic tasks, written in Rust.
Documented measurements
  • DeepSWE v1.1 44.4 According to results reported by Reflection AI, Beam scored 44.4 in this agentic coding evaluation.
  • SWE Bench Pro v2-Hard 77.2 According to results reported by Reflection AI, Beam scored 77.2 in this coding evaluation.
  • SWE Bench Pro v1 65.5 According to results reported by Reflection AI, Beam scored 65.5 in this coding evaluation.
  • Terminal Bench v2.1 80.1 According to results reported by Reflection AI, Beam scored 80.1 in this evaluation of terminal use.
  • SWE Atlas Codebase QnA 34.6 According to results reported by Reflection AI, Beam scored 34.6 in this evaluation of answering questions about code.
According to the sources, it is suitable for
  • The model is designed for coding and agentic tasks involving tool calls.
  • It supports reasoning over text in a context of up to 1 million tokens.
  • It allows self-hosting and further training to meet the needs of a specific team.
Documented limits
  • Image and audio inputs cannot be processed directly; the model accepts only text.
  • Deployment requires substantial memory: all experts remain in memory, and the BF16 checkpoint is approximately 1 TB.

Reflection AI compares Beam with GLM-5.2 and claims that it achieves comparable results on demanding reasoning tasks with inference compute requirements three to four times lower. The TechCrunch article notes that these claims have not been independently verified.

The card summarizes information from the article and any dated corrections, with a link to the original source. It is not our assessment of the model. It does not yet have a dedicated editorial profile. Model selection and other announcements →

What was added since the original report

Verified updates

  1. New verified information

    Beam has been released under the Apache 2.0 license.; Checkpoints are available in BF16, FP8 and NVFP4 formats.; The release is supported by the transformers, SGLang and vLLM libraries.; The model comes with an open harness written in Rust.; The model supports five reasoning levels: low, medium, high, xhigh and max.

    • Beam has been released under the Apache 2.0 license.
    • Checkpoints are available in BF16, FP8 and NVFP4 formats.
    • The release is supported by the transformers, SGLang and vLLM libraries.
    • The model comes with an open harness written in Rust.
    • The model supports five reasoning levels: low, medium, high, xhigh and max.
  2. New verified information

    Beam focuses on coding, logical reasoning and agentic tasks.; According to the company, reinforcement learning ran for more than four weeks on 10 500 Nvidia GB300 GPUs.; The weights are scheduled for release later in the month of the announcement.; The model is to be released under the Apache 2.0 license.

    • Beam focuses on coding, logical reasoning and agentic tasks.
    • According to the company, reinforcement learning ran for more than four weeks on 10 500 Nvidia GB300 GPUs.
    • The weights are scheduled for release later in the month of the announcement.
    • The model is to be released under the Apache 2.0 license.

Two audiences, two different impacts

What this means

01

For individuals

A developer can choose one of five reasoning levels for each task and, depending on its difficulty, look for a suitable balance between response quality, processing time and compute consumption.

What to do During testing, compare the default medium level with a higher level on the same coding task, and monitor both response quality and response time.
More practical updates →
02

For a business

For business deployments, the model offers the option to adapt it to specific tasks, but the infrastructure budget must account for keeping all 501 billion parameters in memory. The number of active parameters alone is therefore insufficient to estimate operating costs.

Development
What to decide Before a pilot deployment, verify the memory requirements of the chosen checkpoint and the required context length on the planned infrastructure.
More business impacts →
Apache 2.0 Beam mixture-of-experts MoE NVIDIA Reflection Reflection AI SGLang vLLM

Check the original

Event sources

confirmed by 2 independent sources · 3 publishers, 2 independent. We count feeds from the same owner only once.

5
TechCrunch AI independent context · first detected Reflection debuts Beam, an open-weight AI model to rival Chinese models at lower compute cost The Decoder (daily AI news) independent context Reflection's Beam becomes the most capable open-weight model built outside China AI Business independent context Reflection’s Beam model signals deepening split in global AI market AI Business independent context Western open-weight campaign could give enterprises more control over AI An overview of multiple AI topics; AI Radar covers only this event. Hugging Face Blog primary source Welcome Beam: a 501B open model you can own, customize, and scale