# Beam (Reflection)

> Source: https://aiwiki.ai/wiki/beam_reflection
> Updated: 2026-10-06
> Fact-checked: 2026-10-06
> Categories: AI Models, Large Language Models, Mixture of Experts, Reasoning Models
> License: CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/) - attribute to "AI Wiki (aiwiki.ai)"
> Cite as: AI Wiki. "Beam (Reflection)." aiwiki.ai, 6 Oct 2026. https://aiwiki.ai/wiki/beam_reflection
> From AI Wiki (https://aiwiki.ai), the free encyclopedia of artificial intelligence. Reuse freely with attribution.

**Beam** is a text-only [large language model](https://aiwiki.ai/wiki/large_language_model) announced by [Reflection AI](https://aiwiki.ai/wiki/reflection_ai) on October 5, 2026 for coding, reasoning and agentic tasks.[1]

## Preview and planned weights

On October 7, 2026, Beam was a selective preview. Reflection planned to release its [weights](https://aiwiki.ai/wiki/open_weights), technical report and model card later in October, with Apache 2.0 licensing for the weights.[1]

## Published specifications and beta limits

| Property | Published value |
| --- | --- |
| Architecture | Sparse [mixture of experts](https://aiwiki.ai/wiki/mixture_of_experts)[1] |
| Total parameters | 501 billion[1] |
| Active parameters | 23 billion[1] |
| Pretraining tokens | 23.8 trillion, reported by Reflection[1] |
| API model identifier | `Beam-501B-A23B`[2] |
| Knowledge cutoff | June 30, 2026[2] |
| Beta context window | 256K tokens, counting prompt and generated output together[2] |
| Maximum output | 128K tokens[2] |

The developer documentation warns that the context window may change during beta.[2]

## Training and expert routing

Reflection reports a four-week [reinforcement-learning](https://aiwiki.ai/wiki/reinforcement_learning) run using 10,500 NVIDIA GB300 GPUs and more than 100 million rollouts.[1]

Beam's expert balancing builds on auxiliary-loss-free balancing, with cosine-decayed expert-bias updates.[1]

The preceding DeepSeek-V3 method adds an adjustable bias to each expert's routing score. Recent loads determine bias updates; the bias changes expert selection, rather than the value that weights an expert's output.[8]

## API use

The Reflection API is in beta, with access opening through a waitlist. Its OpenAI-compatible base URL is `https://api.reflection.ai/openai/v1`, supporting Chat Completions and Models. This is Reflection's service, rather than an OpenAI-hosted model.[4]

Reflection documents Beam configurations for Mirror CLI, Pi, [OpenCode](https://aiwiki.ai/wiki/opencode) and Hermes. Mirror CLI uses Beam by default; the other integrations use its model identifier, the compatible endpoint and a Reflection API key.[5]

### Reasoning controls

`Beam-501B-A23B` always reasons. Its `reasoning_effort` values are `low`, `medium`, `high`, `xhigh` and `max`; omitting the setting selects `medium`. Reasoning cannot be disabled through this setting.[3]

The response separates `message.reasoning_content` from the final answer in `message.content`. Reasoning tokens count toward the completion budget. If that budget runs out during reasoning, `finish_reason` can be `length` and the answer can be null.[3]

### Tool calls and structured output

With [function calling](https://aiwiki.ai/wiki/function_calling), Beam requests application-defined functions; the application executes them and returns results. Its generated JSON arguments require validation before use. The conversation must retain the assistant's tool-call message, including reasoning, followed by one tool-result message for each call.[6]

[Structured output](https://aiwiki.ai/wiki/structured_output) uses `response_format`: `json_object` requests a JSON object, while `json_schema` supplies a schema. Strict schema mode requires every object to set `additionalProperties: false` and require all its properties. Applications must check for refusals and truncated output before parsing the answer.[7]

## Vendor-reported evaluation

| Benchmark | Reflection's reported score |
| --- | --- |
| [SWE-bench](https://aiwiki.ai/wiki/swe_bench) Verified | 80.9[1] |
| [Terminal-Bench](https://aiwiki.ai/wiki/terminal_bench) v2.1 | 80.1[1] |

Reflection estimates generation compute as `2 × active parameters × mean generated tokens per attempt`, including reasoning and answers. The estimate excludes prompt prefill, context-dependent attention and serving overhead. It is an approximate compute comparison, not measured inference cost.[1]

## References

1. Reflection. ["Introducing Beam: Reflection's 501B open-weight model"](https://reflection.ai/blog/introducing-beam). October 5, 2026.
2. Reflection Developer Docs. ["Models"](https://developers.reflection.ai/models). Accessed October 7, 2026.
3. Reflection Developer Docs. ["Reasoning"](https://developers.reflection.ai/reasoning). Accessed October 7, 2026.
4. Reflection Developer Docs. ["Introduction"](https://developers.reflection.ai/). Accessed October 7, 2026.
5. Reflection Developer Docs. ["Coding agent quickstart"](https://developers.reflection.ai/coding-agents-quickstart). Accessed October 7, 2026.
6. Reflection Developer Docs. ["Tool calling"](https://developers.reflection.ai/tool-calling). Accessed October 7, 2026.
7. Reflection Developer Docs. ["Structured outputs"](https://developers.reflection.ai/structured-outputs). Accessed October 7, 2026.
8. DeepSeek-AI et al. ["DeepSeek-V3 Technical Report," section 2.1.2](https://arxiv.org/html/2412.19437v2#S2.SS1.SSS2). arXiv:2412.19437v2. February 18, 2025.