Reflection Beam

Open frontier model for agentic coding and long-horizon work, efficient by design

About model

Beam is Reflection's first open model: a Mixture-of-Experts, decoder-only transformer with roughly 501B total parameters and 23B active per token, built for the two workloads that lead frontier enterprise use, writing and maintaining code and executing long-horizon, multi-step tasks with tools. Beam is text-in, text-out by design: these workloads run on code, tool calls, logs, documents, and structured outputs, and every training decision targets depth in those domains. Reflection reports particularly strong reasoning efficiency, reaching solutions with fewer reasoning tokens and translating into higher throughput per GPU-hour. The model supports a context window of approximately 1M tokens, works out of the box with the runtimes, frameworks, and coding harnesses developers already use, and ships under the Apache 2.0 license. Beam is the first in a family, establishing the training recipe and developer surface that subsequent, larger Reflection models will build on. Available on Together AI.

‍

Total Parameters (23B active)

501B

MoE routing built for efficient frontier-scale inference

Context Window

Entire repositories and long agent trajectories in a single session

Open License

Apache 2.0

Fully open license for commercial use, modification, and redistribution

Model key capabilities

Agentic Coding & Software Engineering: Solves real GitHub issues, generates code, and handles CLI and DevOps work within the coding harnesses developers already use
Long-Horizon Tool Use: Executes multi-step workflows with web search and tools, built for long-horizon task completion rather than single-shot calls
Efficient Dense Reasoning: Reaches solutions with fewer reasoning tokens; Reflection reports compute efficiency several times that of comparable open models, meaning higher throughput per GPU-hour
Production-Ready Infrastructure: 99.9% SLA, available on serverless and dedicated infrastructure

API usage

cURL
Python
Typescript

Endpoint:

reflection/Beam-1.0-501B-A23B

curl -X POST "https://api.together.xyz/v1/chat/completions" \
  -H "Authorization: Bearer $TOGETHER_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "reflection/Beam-1.0-501B-A23B",
    "messages": [
      {
        "role": "user",
        "content": "What are some fun things to do in New York?"
      }
    ]
}'

from together import Together

client = Together()

response = client.chat.completions.create(
  model="reflection/Beam-1.0-501B-A23B",
  messages=[
    {
      "role": "user",
      "content": "What are some fun things to do in New York?"
    }
  ]
)
print(response.choices[0].message.content)

import Together from 'together-ai';
const together = new Together();

const completion = await together.chat.completions.create({
  model: 'reflection/Beam-1.0-501B-A23B',
  messages: [
    {
      role: 'user',
      content: 'What are some fun things to do in New York?'
     }
  ],
});

console.log(completion.choices[0].message.content);

Model card
Architecture Overview:
• Mixture-of-Experts, decoder-only transformer with approximately 501B total parameters and 23B active per token
• Context window of approximately 1M tokens
• Text input and text output, with English-focused alignment; the model retains broader language exposure from pre-training diversity
• Released in BF16, FP8, and NVFP4 variants with speculative decoding support

Training Methodology:
• Every training decision targeted depth in text-native enterprise workloads: code, tool calls, logs, documents, and structured outputs
• Ships with an open-source agentic harness so Beam's agentic behavior reproduces consistently across deployments, alongside post-training and RL recipes
• The first model in a family: Beam establishes the training recipe, developer surface, and integrations that subsequent, larger Reflection models will build on

Performance Characteristics:
• Reflection positions Beam at the frontier of Western open models in the 500B-parameter class across reasoning, coding, and agentic benchmarks
• Reflection reports reasoning efficiency well ahead of comparable open models, reaching solutions with fewer tokens at lower inference-time compute
• Compatible with Claude Code, Hermes, OpenCode, Pi coding harnesses, and Reflection's own Mirror agentic harness
• Benchmark tables arrive with the launch materials
‍
Prompting
Together AI API Access:
• Access Reflection Beam via Together AI APIs using the endpoint reflection/Beam
• Authenticate using your Together AI API key in request headers
• Plugs directly into major coding harnesses for agentic software engineering workflows
• English text in, text out, across a context window of approximately 1M tokens
• Available on Together AI serverless and dedicated infrastructure
‍
Applications & use cases
Agentic Software Engineering:
• Resolve real GitHub issues and maintain codebases end to end
• Run CLI-driven DevOps work through agentic sessions
• Hold entire repositories in the million-token window across long tasks

Long-Horizon Agent Workflows:
• Execute multi-step workflows combining web search, tools, and structured outputs
• Sustain long trajectories at high throughput with 23B active parameters per token
• Convert chat inputs into completed multi-step productivity tasks

Math & Science Reasoning:
• Solve graduate-level science problems and applied mathematics
• Reason through theorems with dense, token-efficient chains of thought
• Serve reasoning workloads where token cost and latency budgets matter
‍