Beam's weights are out: Reflection ships its 501B open-weight model under Apache 2.0

• Reflection AI released Beam's weights on October 16, 2026, on Hugging Face under an Apache 2.0 license (repository reflection/Beam-1.0-501B-A23B).
• Beam is a 501-billion-parameter mixture-of-experts model with 23 billion active parameters per token: text-only, with up to a 1-million-token context window. It was introduced on October 5, 2026, in limited early access.
• Training: pre-trained on 23.8 trillion tokens, then refined in a reinforcement-learning run of more than 100 million rollouts on 10,500 NVIDIA GB300 GPUs over four weeks.
• Checkpoints ship in BF16, FP8, and NVFP4 precision formats, plus a speculative-decoding drafter model, fused kernels, and ready-to-use serving-engine configurations.
• More than 40 launch partners form what Reflection calls an “open operating system for agents”; day-0 framework support covers Hugging Face Transformers, vLLM, SGLang, Unsloth, Axolotl, TokenSpeed, and NVIDIA Dynamo. Baseten offers day-0 Model APIs at $1.30 / $4.30 / $0.13 per million input / cache-input / output tokens.
• Reflection-reported benchmarks: 90.5% on GPQA Diamond, 80.9% on SWE-bench Verified, 97.8% on AIME 2026, 80.1% on Terminal-Bench 2.1, and 78.7% on MCP Atlas. The release also open-sources Mirror CLI, Reflection's internal terminal agent harness.

Reflection AI today made the weights of Beam, its first open-weight model, available on Hugging Face under an Apache 2.0 license — delivering the weights drop it promised when Beam was introduced on October 5. The release comes with a model card carrying final benchmarks, a Beam Safety Technical Report, post-training recipes, and a launch ecosystem of more than 40 partners. It is the biggest US open-weights drop since Beam's introduction, and arguably the most permissive: Apache 2.0, no strings attached.

What's in the release

Beam is a 501-billion-parameter mixture-of-experts model with 23 billion active parameters per token: text-only, with a context window of up to one million tokens. The weights ship in BF16, FP8, and NVFP4 precision formats alongside a speculative-decoding drafter model, fused kernels for lower latency and higher throughput, and ready-to-use configurations for popular serving engines. Alongside the weights, Reflection published training recipes — code templates for fine-tuning — plus tools that measure how customizations change the model's capabilities and safety guardrails.

The training story behind it is substantial: Beam was pre-trained on 23.8 trillion tokens, then refined through a reinforcement-learning run that generated more than 100 million rollouts on 10,500 NVIDIA GB300 GPUs over four weeks. Reflection says roughly 1.3 billion sandboxes ran across training, evaluation, and ablations, with its sandbox fleet peaking at 883,006 vCPUs across 23 clusters on GCP and AWS. SiliconANGLE reports the company developed custom reliability-optimization algorithms to train with practically zero cluster downtime — no small feat at that scale. Self-distillation recipes are promised in the coming days.

The ecosystem play

Reflection is framing this as more than a weights drop: an ecosystem of more than 40 partners and integrations that it calls an “open operating system for agents” — the model as the kernel, with reference agentic integrations, deployment containers, inference integrations, and customization pipelines open above it. Launch platforms named by the company include AWS, Baseten, Cloudflare, CoreWeave, Databricks, Fireworks AI, Google Cloud, LM Studio, Microsoft Azure, Modal, Nebius, NVIDIA NIMs, Ollama, OpenRouter, Oracle, Together AI, Vercel, and Weights & Biases; Dell and HPE deliver Beam behind the firewall for on-premises and air-gapped environments. Reflection also opened a public beta of its first-party API, which it says is free to start.

Framework support is day-zero: Hugging Face Transformers, vLLM, SGLang, Unsloth, Axolotl, TokenSpeed, and NVIDIA Dynamo all work out of the box, and AWS offers a managed JumpStart container for deployment. Baseten, which brought Beam to its platform on day one, prices its Model APIs at $1.30 per million input tokens, $4.30 for cache input, and $0.13 for output — a pricing structure that underlines Beam's efficiency pitch.

Mirror CLI, and the benchmarks

Tucked into the release is an unexpected open-source bonus: Mirror CLI, the terminal agent harness Reflection built for its own engineers. It pairs a Python frontend covering prompts, skills, and developer tooling with a Rust agent core, and supports resumable conversations, parallel agents in separate worktrees, project instructions, MCP connections, and Python extensions — with integrations into Slack, Linear, and GitHub carried over from internal use.

On benchmarks, Reflection reports 90.5% on GPQA Diamond, 80.9% on SWE-bench Verified, 97.8% on AIME 2026, 80.1% on Terminal-Bench 2.1, and 78.7% on MCP Atlas. Those are company-reported numbers; independent scores will arrive now that the weights are public. On release day, Artificial Analysis's tracker scored Beam 30 on its Intelligence Index, second among US open-weights models.

Why it matters

Beam's introduction already positioned Reflection as the US answer to China's open-weights leaders; this release makes that challenge downloadable. An Apache 2.0 license on a 501B-parameter frontier model — with the training recipes to fine-tune it and the harness its own engineers used — is the most aggressive open play by a US lab this year. The question now is whether developers actually build the agent ecosystem Reflection is pitching, or whether the weights end up as another very large download.