Beam
An open-weight large language model released by the Reflection team, with a parameter scale of 501 billion (501B). It ranks among the top tier of current open-weight models and supports local deployment, quantization, fine-tuning, and further development.
Core Facts
Beam's total parameter count is 501B, with 23B parameters activated per token (about 4.6% of the network is activated).
For coding agent tools such as Claude Code, Codex, and Cursor, you need to first install the latest version of Vercel CLI and run the 'vercel ai-gateway setup' configuration command.
Open-weight models refer to models whose parameters are publicly released and can be run locally, but whose training data and code are not necessarily open-source; typical examples include Meta's Llama, Mistral, and DeepSeek.
Timeline (last 90 days)
Beam's pricing, coverage, and deployment scale have not been disclosed.
Beam has entered real deployment scenarios rather than remaining at the concept-demo stage.
Beam is used to let deployed soldiers interact in real time with their children.
Beam is a device or service supporting remote presence interaction, whose core purpose is to let people in different locations share a moment.
Beam's specific technical approach (such as a fixed or mobile form factor) cannot be confirmed from public information.
The U.S. Army and the USO provided support for the Beam project.
Beam's specific capability boundaries, the actual performance of Reflection AI's models in the tests, and whether Factory will turn the test conclusions into product features or pricing strategies are all currently undisclosed.
Factory integrated Beam with Reflection AI's models and stress-tested it in real coding and agent workflows.
By Beam's official release day, Factory already clearly understood how the model behaves on the workloads customers care about most.
Factory's testing covered key stages such as tool error recovery, conflict handling, and final repository validation.
40 more timeline events
All Facts (20)
Beam's total parameter count is 501B, with 23B parameters activated per token (about 4.6% of the network is activated).
65%VerifiedFor coding agent tools such as Claude Code, Codex, and Cursor, you need to first install the latest version of Vercel CLI and run the 'vercel ai-gateway setup' configuration command.
65%VerifiedOpen-weight models refer to models whose parameters are publicly released and can be run locally, but whose training data and code are not necessarily open-source; typical examples include Meta's Llama, Mistral, and DeepSeek.
65%VerifiedReflection AI released an open-source model called Beam, with approximately 501 billion total parameters and only 23 billion activated parameters, using a MoE architecture, designed specifically for reasoning, coding, and agentic workflows, trained end-to-end from scratch, with weights to be made available this month.
70%UnverifiedBeam's pricing, coverage, and deployment scale have not been disclosed.
50%UnverifiedThe U.S. Army and the USO provided support for the Beam project.
50%UnverifiedBeam has entered real deployment scenarios rather than remaining at the concept-demo stage.
50%UnverifiedBeam's specific technical approach (such as a fixed or mobile form factor) cannot be confirmed from public information.
50%UnverifiedBeam is a device or service supporting remote presence interaction, whose core purpose is to let people in different locations share a moment.
50%UnverifiedBeam is a model or model service that has not yet been officially released, with early access provided through the Together platform.
50%UnverifiedBeam's specific capability boundaries, the actual performance of Reflection AI's models in the tests, and whether Factory will turn the test conclusions into product features or pricing strategies are all currently undisclosed.
50%UnverifiedBy Beam's official release day, Factory already clearly understood how the model behaves on the workloads customers care about most.
50%UnverifiedThe core signal conveyed by the tweet is how much Factory values pre-release models, rather than an endorsement of the performance of Beam or Reflection AI's models.
50%UnverifiedFactory's testing covered key stages such as tool error recovery, conflict handling, and final repository validation.
50%UnverifiedThe Beam model can read in an entire codebase at once and handle long tasks without splitting.
50%UnverifiedThe Qwen3.8 27B model scores 34, higher than Beam (501B) at 30 and Nemotron 3.5 Bolt (74B) at 21.
50%UnverifiedReflection AI says Beam was trained end-to-end from scratch rather than adapted from an existing open-source model, and the full weights will be released later this month.
50%UnverifiedBeam reaches 81.6% on SWEBench Verified (BF16).
50%UnverifiedBeam reaches 76.5% on SWE Bench Pro v2-Hard (BF16), 80.1% on Terminal Bench v2.1 (BF16), 90.6% on GPQA Diamond (BF16), and 80.8% on MCP Atlas tool calling (BF16).
50%UnverifiedBeam supports five levels of reasoning effort control (from low to max) via the reasoning_effort parameter.
50%