Blog
Reflection releases Beam, a 501B open-weight coding and agent model under Apache 2.0
On 16 October 2026 Reflection released the weights of Beam, a 501B-parameter mixture-of-experts model with 23B active parameters and a 1M-token context, under Apache 2.0. It targets coding and agent work and is available on major clouds. Benchmark results are Reflection's own.
Reflection, a US AI lab, released the weights of Beam, its first open-weight model, on 16 October 2026. The model is on Hugging Face under the Apache 2.0 license, which allows commercial use and modification. Reflection says Beam is built for coding, reasoning and long-running agent tasks.
The model at a glance
According to Reflection's model card and Hugging Face's launch post:
- Architecture: a sparse mixture-of-experts (MoE) model with 501 billion total parameters, of which about 23 billion are active for each token
- Context window: 1 million tokens
- Input and output: text only, developed mainly for English
- Knowledge cutoff: 30 June 2026
- Checkpoints: BF16, FP8 and NVFP4, plus a small "drafter" model for speculative decoding that speeds up generation
- Reasoning settings: five effort levels, from low to max, plus tool calling
In a mixture-of-experts model, only a few "expert" sub-networks run for each token. Beam can therefore generate text at roughly the speed of a much smaller model. All of its weights still have to be held in memory, though.
What it takes to run
This is not a laptop model. Hugging Face's engineers say:
- The full BF16 checkpoint is about 1 TB.
- The NVFP4 checkpoint shrinks that to about 280 GB but needs NVIDIA Blackwell GPUs.
- They suggest planning for four B200 GPUs for practical serving, while FP8 checkpoints suit older Hopper-generation GPUs such as the H200.
Most teams will reach Beam through a hosted provider. Reflection lists launch availability on platforms including AWS, Google Cloud, Microsoft Foundry, Databricks, Cloudflare, Together AI, Fireworks AI, Baseten, OpenRouter and Ollama. For on-premises and air-gapped setups, it says Dell and HPE deliver Beam inside customers' own data centres. Reflection's own API is in public beta and free to start.
Tools around the model
Reflection is releasing more than weights:
- Mirror CLI, an open-source terminal coding harness that Beam was trained alongside. Reflection says it supports parallel agents, approvals for tool calls and extensions.
- Day-zero support in Transformers, vLLM and SGLang, plus fine-tuning and reinforcement learning recipes built with partners including NVIDIA, Hugging Face and RadixArk.
- Reflection says Beam also works with other open harnesses such as OpenCode, Hermes and Pi.
Benchmarks are Reflection's own
Reflection published a long table of benchmark results. On its own figures, the BF16 model scores 81.6% on SWE-bench Verified, 80.1% on Terminal-Bench 2.1 and 90.6% on GPQA Diamond. These are company-reported numbers, and we had not seen independent reproductions at the time of writing. Hugging Face's team also published informal "vibe evals" on five tasks and stressed that they are not benchmarks.
Reflection also published a Beam Safety Technical Report describing its safety methods and evaluations. It thanks outside groups including Gray Swan, Irregular and SecureBio for testing before release.
Why it matters
Large open-weight models give companies an alternative to closed APIs. They can run the model on their own infrastructure, keep data in-house and fine-tune it for their own tasks. Beam joins a busy week for big open releases. It ships under one of the most permissive licenses available, but its size means self-hosting is realistic only for organisations with serious GPU capacity.
If you want to try it:
- Test it free on Reflection's platform or through a hosted provider.
- Read the model card and safety report before deploying it with real data.
- Run your own evaluations on your own tasks rather than relying on published benchmarks.