The news
Reflection AI Beam, the first open-weight model from the Nvidia-backed startup, was unveiled on October 5, 2026. Reflection says the 501-billion-parameter model is competitive with larger Chinese open models such as Z.ai's GLM 5.2 while using less inference compute.
Beam is a sparse mixture-of-experts model, a design in which only part of the network works on each piece of text. Of its 501 billion parameters, 23 billion are active at a time. Reflection said it pretrained Beam on 23.8 trillion tokens from the web and proprietary licensed datasets, in under four weeks on 6,144 Nvidia GB300 NVL72 GPUs, then ran reinforcement learning on 10.5K GB300 GPUs for four weeks, generating more than 100 million rollouts.
By Reflection's own benchmarks, Beam lands close to GLM 5.2, ahead on some tests and behind on others. It scored 80.1 on Terminal Bench v2.1 against GLM 5.2's 81.0, and 65.5 on SWE Bench Pro v1 against 62.1. Kimi K3 scored higher on every test where both were reported, and Reflection acknowledged that Kimi K3 remains ahead on raw capability. GLM-5.2 has roughly 744 billion total parameters with 40 billion active, according to TechCrunch.
Reflection's pitch is efficiency. The company said Beam reaches scores comparable to GLM-5.2 on advanced reasoning benchmarks while using three to four times less inference compute. It also said those compute figures are estimates built from third-party data from Artificial Analysis and DataCurve, excluding prompt processing and serving overhead, and so are not measured inference costs.
Beam is in final red-teaming and evaluation, with early access by waitlist. Reflection said it will release the weights, a technical report, a model card and safety evaluation results later in October under the Apache 2.0 license, with distribution partners and open-source library support. The model is text-only. TechCrunch noted that Reflection's performance claims have not been independently verified.
Founded in 2024 by two former Google DeepMind researchers, Reflection has raised roughly $4.7 billion from backers including Nvidia (NVDA), Sequoia Capital and Lightspeed Venture Partners, and a funding round in April 2026 valued it at $25 billion before the new money, TechCrunch reported. In the summer of 2026, it signed deals worth more than $7 billion with SpaceX and Nebius for access to Nvidia GB300 chips through 2029.
The numbers
- Total parameters
- 501 billion
- Active parameters per token
- 23 billion
- Pretraining data
- 23.8 trillion tokens
- GPUs used for pretraining
- 6,144 Nvidia GB300
- Inference compute vs GLM-5.2 on reasoning
- 3-4x less (company estimate)
- Reflection funding raised
- About $4.7 billion
Why CEOs should care
For CTOs and AI buyers, open weights under Apache 2.0, a permissive license, mean a company can run Beam on its own infrastructure, fine-tune it and avoid per-token API fees. For teams whose legal or security reviews have slowed adoption of Chinese-developed open models, Beam, which Reflection says advances the Western open-weight frontier, may offer another option. The test is whether it holds up on your workloads once the weights are out.
For CFOs, the efficiency claim is the number to probe. A mixture-of-experts model that activates 23 billion parameters per token can be cheaper to serve than larger models, but Reflection's three-to-four-times figure is its own estimate that leaves out prompt processing and serving overhead. Budget owners should ask for measured cost per completed task on their own prompts and hardware before switching.
For CISOs, the safety work is still pending. Reflection says it will publish safety evaluations in its technical report and open-source the tests it used. Before deploying any open-weight model, security teams should decide who can fine-tune copies, how those copies are tracked and what red-team results they need to see first.
The bigger picture
Chinese labs have set the pace in open-weight models, and Reflection's chief executive has previously said the best open models come from China, TechCrunch reported. Beam's benchmark table compares it mainly with GLM, Kimi, Qwen and DeepSeek models, and TechCrunch framed it as a possible Western answer to DeepSeek, Qwen and Z.ai. Nvidia, an investor, also benefits from a strong open ecosystem that runs on its chips: Beam was trained on Nvidia GB300 hardware, and Reflection is testing a sovereign AI factory partnership with Shinsegae Group in South Korea.
What’s next
The key moments come later in October, when Reflection plans to release Beam's weights, technical report and safety results. Independent benchmark runs will show whether the efficiency and capability claims hold, and the list of distribution partners will show how easily companies can deploy it. Reflection said Beam is the first in a series and that it is already training its next model.
What “Fact-checked” means
Fact-checking means testing a story’s facts against the evidence before it is published. This story went through at least two separate checks before this version was published.
- What we checked
- Its names, figures, dates, job titles, quotes and who said what were checked against the story’s sources, including its main source where it could be opened. The headline was checked for accuracy and overstatement.
- How
- A first check reviewed the whole story. If it passed, a second, skeptical check went back to the sources to look for mistakes in the most important facts. If a check flagged the story, it was edited to fix the problems found, and a separate re-check then reviewed the whole story again.
- Who
- The checks are made by our newsroom, as steps kept separate from the writing, under rules set by our editor, Hussein Mukhtar. A story the checks still flag is held for the editor, who decides whether it is fixed, published or dropped.
- If something is wrong
- “Fact-checked” does not mean error-free. If a material error is found after publication, we correct the story and add a note saying what changed. Report an error
Companies in this story
