Reflection AI Beam: Weight Release and Coding Benchmarks

Announced on October 5, downloadable on October 16

Reflection AI announced Beam, its first open-weight model, on Monday, October 5, 2026. On that day the model was in final red-teaming and evaluation, open only to a select group through a waitlist on Reflection's platform, and Reflection promised the weights, a technical report and a model card later in the month. TechCrunch reported the launch the same day and noted that Axios had reported over the weekend that it was close. [1][2]

On October 16 Reflection published the weights on Hugging Face under the Apache 2.0 license, in BF16, FP8 and NVFP4 formats with a speculative-decoding drafter model, alongside the model card and a safety technical report. It opened a free-to-start public beta of its own API, released an open-source terminal harness called Mirror, and listed more than 20 platforms offering it at launch, including Amazon Web Services, Google Cloud, Microsoft Foundry, Fireworks AI, OpenRouter and Together AI, with Dell and HPE for on-premises installs. [3][5]

What Reflection built

Beam is a text-only sparse mixture-of-experts model. The model card lists 501.3 billion total parameters, 23.4 billion active, 52 layers, 256 experts, a 1 million token context window and a June 30, 2026 knowledge cutoff, and says it was developed mainly for English. Reflection pretrained it on 23.8 trillion tokens on 6,144 NVIDIA GB300 GPUs in under four weeks, then ran reinforcement learning on 10,500 GB300 GPUs for four weeks, generating more than 100 million rollouts from a pool of nearly one million coding, agentic and STEM environments. [1][5]

Reflection's pitch is efficiency. It says Beam matches GLM 5.2 on advanced reasoning tests while using three to four times less inference compute, with a larger gap against the 2 trillion parameter class, and offers a reasoning-effort setting to trade length against accuracy. The compute figure is Reflection's estimate. The company, founded in 2024 by two former Google DeepMind researchers, had raised about $4.7 billion according to TechCrunch's reading of PitchBook, including its October 2025 round. [1][2]

Where Reflection's own table puts it

Reflection's launch table compares Beam with two US open models, Thinking Machines' Inkling and NVIDIA's Nemotron 3 Ultra, and with Chinese models including GLM 5.2, GLM 5.3, Kimi K3, Qwen 3.8 Max and DeepSeek V4.1 Flash. Beam leads the US models on every coding row where they report a score, for example 80.1 against Inkling's 63.8 and Nemotron's 56.4 on Terminal-Bench 2.1. It roughly ties GLM 5.2 (80.1 to 81.0 on Terminal-Bench 2.1, 44.4 to 44.0 on DeepSWE v1.1) and trails Kimi K3 and GLM 5.3 on every coding row where they appear, such as 44.4 against 68.0 and 61.0 on DeepSWE. Reflection revised its results on October 8 and the post now carries the revised table, which is the one quoted here. An archived copy of the October 5 post has the same coding scores; the only Beam score that changed in the tables was CritPT, a reasoning test, from 15.6 to 16.3. [1][4]

Reflection said as much itself: Kimi K3 and similar models remain ahead on raw capability, and Beam's advantage is cost at inference time. TechCrunch noted that none of the performance claims had been independently verified and that Inkling, unlike Beam, accepts images. The final model card, published with the weights, reports slightly different numbers for each precision format, for example 43.3 percent on DeepSWE in BF16. [1][2][5]

The fair reading is that Beam moved the best US open-weight results forward without reaching the Chinese leaders, and that its claim rests on efficiency per token rather than top scores. Every number above is Reflection's selection and Reflection's run. [1][2]

Safety work and what was promised

Reflection trained a separate safety and alignment model from the same pretrained checkpoint and merged it with the main reinforcement-learning model by distillation from both teachers. It described three layers of behavior rules, adversarial rounds that fed successful attacks back into training, and safety reinforcement learning aimed at cutting both harmful compliance and needless refusals. At announcement it promised to publish its safety results and to open-source its internal safety evaluations. [1]

The October 16 model card reports that, on Reflection's in-house Iris Safety Evaluations, the BF16 model refused or redirected 88.9 percent of harmful requests and complied with 81.5 percent of harmless ones. Its usage terms explicitly allow distillation, fine-tuning, quantization and training other models on its outputs. [5]