# Beam 1.0

> Beam 1.0 is Reflection's open-weights mixture-of-experts model — 501B total parameters, 23B active across 256 experts — built for token-efficient reasoning, agentic tool use, and software-engineering workloads. Reasoning is always on with a tunable effort level; the Apache-2.0 weights are on Hugging Face.

**ID**: `reflection/beam-1.0`  
**Creator**: reflection  
**Category**: text  
**Context**: 262K tokens  
**Input**: $0  
**Output**: $0  
**Released**: 2026-10-05  
**Web page**: https://anyrouter.dev/model/reflection/beam-1.0

**Input modalities**: text  
**Output modalities**: text  

**Capabilities**: chat, streaming, reasoning, coding, function-calling, structured-output

**Supported parameters**: max_tokens, temperature, top_p, top_k, tools, tool_choice, response_format, reasoning, include_reasoning, reasoning_effort

**Aliases**: `reflection-ai/beam-1.0-501b-a23b`, `reflection/beam-1.0-501b-a23b`

## Usage

**Endpoint**: `POST https://anyrouter.dev/api/v1/chat/completions`

```bash
curl https://anyrouter.dev/api/v1/chat/completions \
  -H "Authorization: Bearer $ANYROUTER_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
  "model": "reflection/beam-1.0",
  "messages": [
    {
      "role": "user",
      "content": "Say hi in 3 words."
    }
  ]
}'
```

## Providers

| Provider | Input | Output |
| --- | --- | --- |
| hermes-agent | $0 | $0 |
| hermes-agent-byok | $0 | $0 |

## Model card

- Architecture: mixture of experts [1]
- Total parameters: 501.3B [1]
- Active parameters: 23.4B [1]
- Native context: 1048576 tokens [1,2]
- Training cutoff: 2026-06-30 [1,2]
- License: Apache-2.0, open weights (https://huggingface.co/reflection/Beam-1.0-501B-A23B) [1]

### Strengths [1]

- Token-efficient reasoning — comparable quality with ~3x fewer reasoning tokens
- Agentic tool use: multi-step workflows, MCP tool integration, browsing
- Strong SWE scores: 81.6% SWE-bench Verified, 80.1% Terminal Bench v2.1

### Known limits [1,2]

- Developed predominantly for English
- Reasoning is mandatory — there is no non-reasoning mode
- Hosts serve a 262144-token window, below the model's 1M training context

### Use cases [1]

- Agentic coding: real GitHub issues, CLI and DevOps work
- Long-horizon tool workflows over MCP
- Math, science, and applied reasoning

### API notes [1,2]

- reasoning_effort accepts low/medium/high/xhigh/max; default max, cannot be turned off
- Chat template is a modified Harmony format
- Tool calling and structured outputs (JSON schema) supported

### Links

- announcement: https://reflection.ai/blog
- model card: https://huggingface.co/reflection/Beam-1.0-501B-A23B
- docs: https://developers.reflection.ai/models

### Sources

1. [reflection/Beam-1.0-501B-A23B model card](https://huggingface.co/reflection/Beam-1.0-501B-A23B/raw/main/README.md) (accessed 2026-10-17)
2. [Reflection API — Models](https://developers.reflection.ai/models) (accessed 2026-10-17)
3. [Cloudflare Workers AI — Beam 1.0 501B A23B](https://developers.cloudflare.com/ai/models/reflection-ai/beam-1.0-501b-a23b/) (accessed 2026-10-17)

## Source

- Upstream docs: https://huggingface.co/reflection/Beam-1.0-501B-A23B