Reflection AI unveils Beam, a 501B open model due under Apache 2.0

Reflection AI announced Beam on 5 October, a mixture-of-experts model with 501 billion parameters of which 23 billion are active. The company's blog post says the weights, a technical report and a model card will follow under an Apache 2.0 licence later this month. Beam is still in final red-teaming and evaluation, so nobody outside Reflection has tested it, and an early-access sign-up is open at platform.reflection.ai.

Reflection says Beam matches Z.ai's GLM 5.2 on reasoning while using three to four times less inference compute, and calls it ahead of other Western open models in its class. Chief executive Misha Laskin told Semafor it is a "workhorse" for coding and agentic tasks, aimed at buyers who "don't or can't use Chinese models." The efficiency claim has not been checked independently.

The blog's tables put Beam at 80.9 on SWE-bench Verified against 77.6 for Inkling, 80.1 on Terminal Bench v2.1 and 97.8 on AIME 2026. They also show it trailing larger rivals: Kimi K3 scores 93.5 against Beam's 90.5 on GPQA Diamond, and 46.9 against 36.2 on Humanity's Last Exam without tools. All the figures are Reflection's own.

On training, Reflection says it pretrained on 23.8 trillion tokens in under four weeks on 6,144 Nvidia GB300 NVL72 GPUs. It then spent four weeks on reinforcement learning across 10,500 GB300s, with more than 100 million rollouts. The post says context was extended to 1 million tokens. TechCrunch describes Beam as text-only.

TechCrunch reports that Reflection, founded in 2024 by two former Google DeepMind researchers, has raised about $4.7 billion at a $25 billion pre-money valuation, with Nvidia, Sequoia and Lightspeed among its backers. Ollama said on X it is "excited" about the model and that more US models are coming to its library. Laskin told Semafor the next model is already in training. Until the weights ship, the claims are Reflection's alone.

https://x.com/ollama/status/2107193986482126883