Beam

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.

Reflection 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.

Beam is offered in three formats—BF16, FP8, and NVFP4—supports deployment via Amazon SageMaker JumpStart, and can also run on Kubernetes clusters running vLLM or SGLang.

Beam reaches frontier level on coding and agent tasks while using about 3-4 times fewer inference tokens than comparable open-source models.

Beam supports five levels of reasoning effort control (from low to max) via the reasoning_effort parameter.

Beam introduces a controllable length penalty into the reinforcement learning reward function, penalizing unnecessary token generation while rewarding successful task completion.

Beam adopts designs such as interleaved local and global attention layers, fine-grained routed experts, and a controlled residual stream to improve numerical stability.

At the end of Beam's pretraining, the busiest expert's load was only 1.04 times the average.

Beam was pretrained on 23.8 trillion high-quality tokens.

Beam combined with DFlash2 speculative decoding achieves 1.72x to 3.03x generation speedup depending on the task and hardware configuration, while preserving the full model's exact output distribution.

Beam 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).

Beam reaches 81.6% on SWEBench Verified (BF16).

Beam's post-training RL training ran for four weeks on 10,500 NVIDIA GB300 GPUs, generating over 100 million rollouts.

Beam is open-sourced under the Apache 2.0 license, with weights hosted on Hugging Face.

Beam is optimized for coding, reasoning, and agent workloads.

Beam has not yet disclosed key information such as benchmarks, license, model weight downloads, training data, context length, and multilingual support.

For developers with limited resources, deploying Beam may require waiting for an official quantized version or adopting a distributed inference solution.

Ollama launched support for Beam as a "Day 0 partner," indicating it is extending from a local model running tool toward a cloud model distribution platform.