1. Home
2. News
3. Open weights

Reflection releases Beam's 501 billion parameter weights under Apache 2.0

Reflection pitches Beam on how few tokens it spends, and the first outside test backs that up but finds it hallucinates often.

By Think Facility · October 16, 2026 · Open weights AI models Save

On October 16, 2026, Reflection put the weights of Beam, its first model, on Hugging Face under an Apache 2.0 license. Reflection says developers can download it, host it themselves and customize it for their own work.

Beam is a mixture-of-experts model with 501 billion parameters, 23 billion of them active at a time. Reflection announced it on October 5 and promised the weights "later this month," so it beat its own deadline by about two weeks.

Size

501 billion total parameters, 23 billion active

License

Apache 2.0

Weights

BF16, FP8, and NVFP4

Context window

1M tokens

Announced

October 5, 2026

What Reflection says it's good at

Reflection built Beam for coding and agent work, and it sells the model on efficiency more than on top scores. In the October 5 post, the company says Beam scores about the same as GLM-5.2 on advanced reasoning tests while using 3 to 4 times less inference compute. It doesn't claim the lead:

The efficiency came from how it was trained. Reflection used "a controllable length penalty that rewards successful solutions while discouraging unnecessary tokens." Early on, scores rose while answers got shorter. Later, as Beam got better at agent tasks, its answers grew again. Users pick the tradeoff themselves (there's a reasoning effort setting for it).

What an outside test found

Artificial Analysis, one of the groups Reflection thanks for testing Beam before release, posted its own scores the same morning, at 9:31 am Eastern. They're more mixed than Reflection's tables, and more specific.

So the token count backs up Reflection's pitch. Beam used about 26,000 output tokens per task, half of what the UAE's K2 Horizon needed for a score one point higher. The hallucination rate cuts the other way: 47%, against 26% and 30% for the two models it was compared with.

Terminal work is weak. Beam scored 3.0% on Terminal-Bench 4.0 in that test. Reflection's own table reports 80.1 on an older version, Terminal Bench v2.1, so I wouldn't set the two numbers side by side.

How big the training run was

Reflection pretrained Beam on 23.8 trillion tokens using 6,144 NVIDIA GB300 GPUs, in under four weeks. The reinforcement learning run after that used 10,500 GB300s for another four weeks and produced more than 100 million rollouts.

Where you can run it

Besides the three weight formats, there's a small drafter model for speculative decoding. vLLM, SGLang and Transformers support Beam from the first day, and Ollama, OpenRouter and the big cloud platforms are on Reflection's list of hosts. Dell and HPE sell it for machines kept behind a company's firewall.

Reflection also open-sourced Mirror CLI, the terminal harness it trained Beam alongside. Recipes for self-distillation, a way to tune Beam on your own tasks, are due "in the coming days." The company says it's already training the next model in the series.