Above: Reflection AI is described by TechCrunch as a Brooklyn-based startup. (Image: Andre Carrotflower, via Wikimedia Commons (CC BY-SA 4.0))
Reflection AI has announced Beam, its first open-weight model. The company's pitch is efficiency: it says Beam reaches reasoning scores comparable to a larger Chinese open model while spending far less compute per answer. The claim comes from Reflection, and the company's own benchmark tables are more mixed than the headline.
What was announced
Beam is a sparse Mixture-of-Experts model with 501 billion total parameters, 23 billion of them active, according to Reflection's announcement. The company says it is built for coding, reasoning and agentic work. It is text-only, has a context length extended to 1M tokens, and was pretrained on 23.8 trillion tokens.
Reflection says the reinforcement-learning stage generated over 100 million rollouts on 10.5K Nvidia GB300 GPUs across four weeks.
When Reflection announced Beam, it had not yet released the model. At launch, Reflection said the model was undergoing final red-teaming and evaluations, with early access by sign-up. It said it would release the weights, technical report, model card and developer artifacts "later this month", with the weights under an Apache 2.0 licence. Reflection's post also says the launch will come with distribution partners and integrations with open-source libraries.
The efficiency claim
The headline figure is in Reflection's post: on advanced reasoning benchmarks, Beam "achieves scores comparable to GLM-5.2 while using 3–4× less inference compute." TechCrunch, which covered the announcement, notes that Reflection's performance claims haven't been independently verified.
Reflection's own footnote also qualifies the number. The compute figures are estimates based on active parameters multiplied by average tokens generated per attempt, using data from Artificial Analysis and DataCurve. Reflection says they exclude prompt prefill, context-dependent attention operations and serving overhead. In its words, they are "an approximate compute comparison rather than measured inference cost."
What the benchmark tables show
Reflection's post says Beam is competitive with larger open models like GLM 5.2 and approaching Qwen 3.8-Max on coding and agentic tasks. It also says that "Where frontier open models like Kimi K3 remain ahead on raw capability, Beam's advantage is efficiency at inference time."
The tables, which Reflection updated on 8 October, put numbers on that. On the four reasoning benchmarks where both models report a score, GLM-5.2 is ahead of Beam:
- AIME 2026: Beam 97.8, GLM-5.2 99.2
- HLE (no tools): Beam 36.2, GLM-5.2 40.5
- CriPT AA: Beam 16.3, GLM-5.2 20.9
- GPQA Diamond: Beam 90.5, GLM-5.2 91.2
"Comparable" is Reflection's word. The gaps are small on AIME and GPQA Diamond and larger on HLE and CriPT AA. On Terminal Bench v2.1, Beam scores 80.1 against 81.0 for GLM-5.2 and 88.3 for Kimi K3.
The comparison set is not only Chinese models. The tables also include Inkling and Nemotron 3 Ultra. TechCrunch reports that Reflection's benchmarks show Beam outscoring Inkling, the open model from Thinking Machines Lab, on four coding tests where both report results. It adds that Inkling is multimodal and Beam is text-only.
Reflection's post says that "Beam advances the Western open-weight frontier." TechCrunch frames the launch as one that could heat up the race to build a Western answer to DeepSeek, Qwen and Z.ai. The "Western answer" phrasing is TechCrunch's; Reflection's own post compares Beam with models from those labs, including GLM 5.2, Qwen 3.8-Max and DeepSeek V4.1 Flash.
The parameter comparison
TechCrunch reports that Z.ai's GLM-5.2 has roughly 744 billion total parameters with 40 billion active. That figure is TechCrunch's. If it holds, Beam is smaller on both counts, which is consistent with Reflection's focus on inference cost.
The company behind it
TechCrunch reports that Reflection is two years old and was founded in 2024 by two former Google DeepMind researchers. Citing PitchBook, it says the company has raised roughly $4.7 billion, with Nvidia among its backers. It also reports that Reflection's last round valued it at $25 billion pre-money. TechCrunch reports, too, that Reflection signed deals worth more than $7 billion in total with SpaceX and Nebius for access to Nvidia GB300 chips through 2029.
TechCrunch adds that Reflection is aiming Beam and future models at enterprises and sovereign nations. It says Reflection did not respond in time to its requests for more information.
What to watch
Three things separate "announced" from "shown" here. The weights have to ship under the licence Reflection described. Independent testers have to reproduce the benchmark scores. And someone outside Reflection has to measure Beam's real serving cost, because the 3–4× figure is the company's own estimate.
Sources
Reflection (2026) 'Introducing Beam: Reflection's 501B open-weight model', Reflection, 5 October 2026, updated 8 October 2026, available at: https://reflection.ai/blog/introducing-beam (Accessed: 10 October 2026).
TechCrunch (Bellan, R.) (2026) 'Reflection debuts Beam, an open-weight AI model to rival Chinese models at lower compute cost', TechCrunch, 5 October 2026, available at: https://techcrunch.com/2026/10/05/reflection-debuts-beam-a-open-weight-ai-model-to-rival-chinese-models-at-lower-compute-cost/ (Accessed: 10 October 2026).