Key points
- Reflection unveiled Beam, with open weights due later this month
- Its own benchmarks show newer Chinese models ahead
- Reflection is selling Beam on efficiency
Reflection, the AI startup backed by Nvidia (NVDA), unveiled its first open-weight model, Beam, on Oct. 5. In its announcement, Reflection says Beam delivers comparable advanced-reasoning scores to Z.ai's GLM-5.2 while using roughly one-quarter to one-third as much inference compute.
Beam is still "undergoing final red-teaming and evaluations," and an early version is open to a waitlist. Reflection said it will release the open weights under an Apache 2.0 license later this month, along with a technical report and model card.
The same announcement shows newer Chinese open models, including Z.ai's GLM-5.3 and Moonshot AI's Kimi K3, ahead of Beam on most tests where both report scores. "Where frontier open models like Kimi K3 remain ahead on raw capability, Beam's advantage is efficiency at inference time," the company wrote. These are company-reported results, and they haven't been independently verified, TechCrunch reported.
| Benchmark | Beam | GLM-5.2 | GLM-5.3 | Kimi K3 | Inkling |
|---|---|---|---|---|---|
| Terminal Bench v2.1 | 80.1 | 81.0 | 88.2 | 88.3 | 63.8 |
| Humanity's Last Exam (no tools) | 36.2 | 40.5 | 42.3 | 46.9 | 29.7 |
| GPQA Diamond | 90.5 | 91.2 | 91.7 | 93.5 | 87.2 |
| SWE-Bench Pro v1 | 65.5 | 62.1 | n/a | n/a | 54.3 |
Scores as reported by Reflection on Oct. 5. "n/a" means no score was reported.
Across the announcement's broader benchmark results, Beam outscores Thinking Machines Lab's Inkling on four coding tests where both report scores, TechCrunch noted. Inkling is multimodal, while Beam handles only text.
Beam has 501 billion parameters
Beam is a mixture-of-experts model with 501 billion total parameters, 23 billion of which are active at a time. Reflection pretrained it on 23.8 trillion tokens in under four weeks on 6,144 Nvidia GB300 GPUs, and it has a 1 million-token context window.
Its reinforcement learning run used 10,500 GB300 GPUs for four weeks and generated more than 100 million rollouts, the company said.
Reflection is selling AI factories
Reflection is aiming Beam at companies, governments and developers that want AI systems they control and can train on their own data. It calls that product an AI factory. In March, it signed a memorandum of understanding with South Korea's Shinsegae Group to build a 250-megawatt AI factory in Korea.
The startup has raised about $4.7 billion from backers including Nvidia, Sequoia Capital and Lightspeed Venture Partners, according to PitchBook data cited by TechCrunch. Its last round valued it at $25 billion before the new money.
Reflection's compute deals with SpaceX (SPCX) and Nebius (NBIS), worth more than $7 billion combined, give it access to Nvidia GB300 chips through 2029, TechCrunch reported.
