Reflection Beam open weights land under Apache 2.0

Reflection released the weights of Beam, its 501-billion-parameter AI model, on Hugging Face on October 16, 2026, under the Apache 2.0 license. Apache 2.0 lets anyone use, change and sell products built on the model with few conditions. That makes Beam one of the largest models available with no custom license terms, and its compressed versions fit on a single eight-GPU server.

The release keeps the promise Reflection made when it announced Beam on October 5, when the weights were due "later this month." "Beam is our first open-weight model," the company wrote in its release post.

What Reflection released

Beam is a sparse mixture-of-experts (MoE) model. Its parameters are split into many smaller expert networks, and only a few run for each token, a word or part of a word. The Hugging Face model card lists the details:

Spec	Beam-1.0-501B-A23B
Total parameters	501.3B
Active per token	23.4B
Experts and layers	256 experts, 52 layers
Context window	1M tokens
Input	Text only, mainly English
Weight formats	BF16, FP8 and NVFP4
License	Apache 2.0

The card describes this as the instruction-tuned model, the version trained to follow requests. Reflection also shipped a smaller draft model for speculative decoding, a trick that speeds up text generation. It added fused kernels, ready-made settings for several serving programs and a Beam Safety Technical Report. A full technical report is still listed as coming. Recipes for self-distillation will follow "in the coming days," SiliconANGLE reports.

Reflection also released Mirror CLI, a coding agent that runs in a terminal, VentureBeat reports. The company calls the whole package an "open operating system for agents."

How the compressed versions score

Quantization stores a model's numbers with fewer bits, which cuts memory at some cost in accuracy. On Reflection's own tests, that cost is small:

Benchmark	BF16	FP8	NVFP4
SWE-bench Verified	81.6	80.2	80.7
SWE-Bench Pro v1	66.1	65.9	65.6
Terminal-Bench 2.1	80.1	78.9	78.4
GPQA Diamond	90.6	90.3	89.3
LiveCodeBench	88.3	88.1	87.4

The 8-bit and 4-bit versions fit on one server with eight H100, H200 or GB300 GPUs, VentureBeat reports. The full 16-bit version needs "tens of GPUs."

Against rivals, Beam is not the top scorer. The chart shows the Terminal-Bench 2.1 figures VentureBeat compiled. DeepSeek's V4.1 Flash leads at 90.6, ahead of Beam's 80.1. Beam beats Meta's Muse Spark Open, which scored 76.4.

The pitch: fewer tokens

Reflection's case rests on efficiency, not raw scores. It claims Beam matches GLM-5.2 on advanced reasoning while using three to four times less inference compute, not counting prompt processing. The model is trained to reach answers with fewer reasoning tokens.

"You don't give up anything. You just penalize it in terms of how many tokens it can use," Ioannis Antonoglou, Reflection's co-founder and CTO, told VentureBeat. VentureBeat estimates that Muse Spark Open uses about 1.8 times Beam's compute per token.

Reflection has not set a price. Its platform beta is "free to start," the company says, and it lists more than 40 launch partners. Dell and HPE will install Beam on customers' own hardware.

The release follows a Financial Times report on October 10 that Nvidia is in early talks to buy Reflection or invest more. Antonoglou called that "baseless rumors" without confirming or denying talks, VentureBeat reports. He described Nvidia as "a close partner and investor."

What this means for developers

• Start with FP8 or NVFP4. Reflection's numbers show the compressed versions within about two points of full precision. They fit on one eight-GPU server instead of a multi-server cluster.
• Count tokens per task, not price per token. Reflection's claim is that Beam reaches answers with fewer tokens. Measure total tokens and time on your own tasks to test it, because that is where any saving appears.
• The license is the simple part. Apache 2.0 carries no revenue caps or custom terms. Legal review should be shorter than for models under custom licenses, such as Muse Spark Open's.
• Treat the scores as vendor numbers. Every benchmark here comes from Reflection or a model card. Independent tests can now start because the weights are public.
• Watch the follow-ups. The full technical report and the self-distillation recipes are still to come. So is any outcome of the reported Nvidia talks, which could change who controls the model's roadmap.