A 501B open model you can own: Reflection’s Beam ships under Apache 2.0

US lab Reflection has released Beam, its first model, with open weights under the Apache 2.0 license. According to the Hugging Face announcement, it is a 501B-parameter mixture-of-experts model that activates 23B parameters per token and handles a 1M-token context.

What is new in Beam

Beam is a text-in, text-out model aimed at coding and agentic work. The announcement says it is built for teams that want frontier-scale capability on their own hardware, infrastructure or cloud. It ships with an open-source custom harness written in Rust.

The architecture is a sparse mixture of experts with 52 decoder layers and a hidden size of 5,120. The first three layers are dense, and the remaining 49 are expert layers with 256 experts each, of which 8 are routed per token. Experts make up 98.4% of the parameters, and there is no shared expert.

Attention is built to keep long contexts affordable. Most layers use a 512-token sliding window, while every sixth layer and the last layer attend globally (43 sliding-window layers and 9 global ones). Reflection says this brings the key-value cache to about 36 GiB per 1M-token sequence in BF16, against roughly 208 GiB if every layer were global.

Beam also offers five reasoning effort levels: low, medium (the default), high, xhigh and max. It can keep its reasoning in context across turns and has a dedicated grammar for tool calling.

Checkpoints, hardware and availability

Three checkpoints are published. The BF16 version is about 1 TB. An FP8 version targets Hopper and Ampere GPUs. An NVFP4 version of about 280 GB targets Nvidia Blackwell. A speculative-decoding drafter called DFlash2 is included.

For the NVFP4 build, the post recommends four B200 GPUs (720 GB). Two B200s can hold the weights but leave little room for activations and cache. Other options listed are a full 8xB200 node, two B300s or 8xRTX Pro 6000. Outside Blackwell, the post points to FP8, for example on eight H200s with SGLang.

Reflection says it supports Transformers (version 5.20.1 or later), vLLM and SGLang from day zero. Hosted access runs through Hugging Face Inference Providers, currently Baseten, Fireworks and Together, with more to come. A one-click Inference Endpoints recipe is also offered. The post lists no per-token or endpoint pricing.

Reported results

The benchmark figures come from Reflection itself and were not independently verified in the post. For the BF16, FP8 and NVFP4 checkpoints respectively, it reports:

• SWE-Bench Verified: 81.6%, 80.2% and 80.7%
• Terminal Bench v2.1: 80.1%, 78.9% and 78.2%
• AIME 2026: 97.7%, 98.0% and 97.5%
• GPQA Diamond: 90.6%, 90.3% and 89.3%
• LiveCodeBench: 88.3%, 88.1% and 87.4%

The quantized versions stay close to the full-precision scores, and on SWE-Bench Pro v2-Hard they come out slightly ahead (78.4% and 78.6% against 76.5%). HLE with tools scores 51.1% on all three.

Fine-tuning examples

The post also covers post-training. With LoRA supervised fine-tuning on a multilingual reasoning dataset, Beam reasoned in the requested language on 16 of 16 held-out prompts after one epoch, against 3 of 16 for the base model. In reinforcement learning with GRPO on math data, full-parameter training lifted AIME25 accuracy from 46.25% to 55.8% at a 4,000-token limit. LoRA GRPO raised it from 42.9% to 64.2% while shortening outputs from 3,188 to 2,197 tokens.

In a smaller demonstration, a fine-tuned Qwen3.5-0.8B model for incident triage reached 76.25% accuracy against 4.69% for a prompted zero-shot setup, and cut latency from 1,218 ms to about 103 ms.

Why it matters

Beam’s economics rest on a bet that is easy to state and costly to test: that a US lab can give away frontier-scale weights and still build a business. Reflection raised $2 billion in a round led by Nvidia, according to AI Business, at a reported valuation of $8 billion. Apache 2.0 leaves little room to charge for the weights themselves. Revenue has to come from hosting, support and enterprise deployments, and the Hugging Face post lists no pricing for any of them.

The company’s main cost argument is inference. Coverage of the launch says Reflection describes Beam as competitive with the larger open model GLM 5.2 on coding and agentic tasks while using 3-4x less inference compute. It also says Reflection concedes that Kimi K3 remains ahead on raw capability. Both are the company’s own claims and have not been independently verified. If they hold, the savings come from the 23B active parameters and the cheap long-context cache, not from the license.

For buyers, the choice is between renting access from Baseten, Fireworks or Together and buying Blackwell hardware. At 280 GB for the NVFP4 build, that hardware means four B200s. Self-hosting only pays off for teams with steady, high-volume agent workloads or strict data-residency rules. Everyone else will probably compare per-token prices once providers publish them, and the 3-4x efficiency claim will matter most at that point.

What’s next

Reflection says it will award $10,000 in Hub credits to the best projects built with Beam. Entries are shared on the model’s discussion page, and the post says more inference providers will be added.