Intel / article

Reflection releases Beam, a 501B open-weight coding and agent model

Source: reflection.ai
Author: Reflection
Date: Published Oct 16, 2026

Context

Reflection has published the weights of Beam, its first open-weight model, on Hugging Face under the Apache 2.0 license (reflection.ai). Beam is a sparse model with 501 billion parameters in total, of which 23 billion are active for any given token, aimed at coding, reasoning and long-running agent work. The weights come in BF16, FP8 and NVFP4 precisions, with a smaller drafter model for , a technique that speeds up generation by having a small model propose tokens that the large one checks. Reflection first introduced Beam on October 5; this post covers the weights release and the tooling and partners around it.

Reflection calls the launch an open operating system for agents: the model is the kernel, and the layers above it are meant to be open too. The most concrete piece is Mirror CLI, the terminal coding Reflection says it trained Beam alongside and is now open-sourcing. Mirror pairs a Python frontend for prompts, skills and tooling with a minimal Rust core that drives the model, which Reflection says also makes it a clean surface for evaluation. It supports resumable conversations, parallel agents in separate worktrees, MCP connections and approval of tool calls. Reflection says Beam has also been tested with other open harnesses, including Pi, OpenCode and Hermes from Nous Research.

The rest of the stack is partnerships. Reflection lists day-zero support in Transformers, vLLM, SGLang, Unsloth, Axolotl, TokenSpeed and NVIDIA Dynamo, fine-tuning and reinforcement-learning workflows built with partners including NVIDIA and Hugging Face, and hosted access across clouds and inference providers, with Dell and HPE offering on-premises deployments. Its X thread says reproducible evaluation recipes are still to come, as are the self- recipes the post promises in the coming days.

Key takeaways · AI-distilled

Beam is a sparse Mixture-of-Experts model with 501B total and 23B active parameters, released under Apache 2.0 with weights in BF16, FP8 and NVFP4 plus a speculative-decoding drafter model, per Reflection.
Reflection says it trained Beam together with its open-source Mirror CLI harness, which pairs a Python frontend for prompts and skills with a minimal Rust agent core that drives the model and doubles as an evaluation surface.
Beam launched with day-zero serving support in Transformers, vLLM, SGLang, Unsloth, Axolotl and NVIDIA Dynamo, and Reflection says it was also tested in the Pi, OpenCode and Hermes harnesses.
Reflection reports running roughly 1.3 billion sandboxes across RL training, evaluation and ablations for its largest Beam runs, and says it will adopt the OpenEnv environment standard for future runs.
Post-training is supported through NVIDIA NeMo RL, Hugging Face TRL, Anyscale SkyRL, RadixArk Miles and Axolotl workflows; Reflection's self-distillation recipes are promised for the coming days but not yet released.

Terms in this piece · Glossary

mixture-of-experts — A model built from many specialist sub-networks where only a few activate per token, giving big-model capability at small-model running cost.
speculative decoding — A speed trick where a small model drafts several tokens ahead and the big model verifies them in one pass, often doubling generation speed.
agent harness — The scaffolding around a model that turns it into a working agent — the loop, the tools it can call, and the rules for when to stop.
distillation — Training a small, cheap model to imitate a big one's outputs, keeping much of the capability at a fraction of the cost.

Read the source reflection.ai