Reflection Beam: Apache 2.0 open weights and an open operating system for agents

Introduction

One day after Meta’s Muse Spark Open weights, Reflection shipped Beam as a fully Apache 2.0 coding and agent MoE. The license difference matters as much as the parameter count. Spark Open is a large multimodal open-weight option under Meta’s Muse Model License. Beam is text-only, smaller at the active-parameter layer, and explicitly pitched with an open agent harness and customization stack.

Primary sources: Beam is now available, the Hugging Face Beam guide, Mirror CLI overview, and Mirror MCP connections. Adjacent: Muse Spark Open routing, Kolibri-1 EU routing, Mistral Large 4, and HydraFusion Max.

What shipped

Reflection’s earlier technical preview described large-scale RL (on the order of 100M+ rollouts) and midtraining that extends effective context to 1M tokens. The October 16 availability post focuses on the open stack: inference formats, partner distribution, Mirror CLI, fine-tuning and RL recipes with NVIDIA, Hugging Face TRL, Anyscale SkyRL, Axolotl, and related tools, plus on-prem paths with Dell and HPE.

Open operating system for agents

Reflection’s framing is useful for architecture conversations even when teams keep another model as the planner. Weights alone do not make an agent system. Beam’s launch bundles:

• Open weights under Apache 2.0 for self-host, VPC, and sovereign deployments.
• Reference agent integrations and Mirror CLI with MCP connections for external tools.
• Deployment containers and day-zero inference integrations to reduce custom serving glue.
• Customization pipelines (SFT, RL, planned self-distillation recipes) so domains can adapt the kernel.

Mirror CLI and MCP

Mirror CLI runs in the project directory, defaults to Reflection’s Beam model when `REFLECTION_API_KEY` is set, and supports manual approval (default) or auto mode for tool calls. MCP servers can be connected via built-in services or JSON definitions under `~/.mirror/mcp/`, so a cloned repo cannot inject servers. That layout is a practical governance detail for enterprises that already standardize MCP configs outside the working tree.

Other coding agents can call Beam through Reflection's OpenAI-compatible endpoint (see the coding agents quickstart) with model ID `Beam-1.0-501B-A23B`. That path matters when a fleet is already standardized on OpenCode or Hermes and only needs a new open-weight backend.

Where Beam sits on the routing board

What this means for developers

• Pull BF16, FP8, or NVFP4 from Hugging Face and pin a commit hash in the model registry.
• Start with a hosted Reflection or partner endpoint if GPU capacity is not ready; keep the same source ID for later self-host cutover.
• Pilot Mirror CLI or an existing OpenAI-compatible harness with MCP allowlists identical to production.
• Re-run SWE/terminal/MCP evals against Muse Spark Open and the current closed planner baseline.
• Do not assume Apache 2.0 removes safety or red-team work; review Reflection’s model card and safety materials before customer data.

What this means for businesses

• Apache 2.0 simplifies commercial legal review relative to limited commercial open licenses, but still needs a formal license and safety packet.
• Partner and on-prem paths (including Dell/HPE messaging) make Beam a candidate for sovereign or air-gapped agent stacks.
• FinOps should compare partner token prices against GPU amortization on the same completed-task metric used for Spark Open.
• Treat Mirror MCP connections as an enterprise connector surface: approve servers outside the repo tree and audit approvals.

Adoption checklist

1. 01
   Classify the use case

   Self-host residency, partner API, or keep Spark Open / closed planners for multimodal or merge-critical work.
2. 02
   Legal pass on Apache 2.0 + safety materials

   Confirm redistribution, modification, and customer-data use before production traffic.
3. 03
   Stand up one inference path

   Partner endpoint or internal vLLM/SGLang with observability and spend caps.
4. 04
   Run three agent evals

   Long refactor, MCP tool loop, and terminal/SWE task against Spark Open and Sol/Opus baselines.
5. 05
   Decide promotion

   Promote only where Beam wins cost-per-completed-task and policy review inside the chosen harness.

Conclusion

Beam gives enterprises an Apache 2.0 US open-weight agent MoE with a serious harness story the day weights land. The technical story is efficient active-parameter inference with a 1M context for coding agents. The business story is ownership and customization without a limited commercial open license. Route it as a deliberate self-host lane next to Muse Spark Open, not as an automatic replacement for every frontier planner.

Sources: reflection.ai — beam is available ; Hugging Face ; developers.reflection.ai — mirror overview ; developers.reflection.ai — mirror mcp ; GitHub ; developers.reflection.ai — coding agents quickstart