Reflection Releases Beam's 501B Open Weights Under Apache 2.0

Reflection just handed out the weights to its flagship model. Beam carries 501 billion parameters, runs agentic coding tasks, and ships under a license that lets companies host it in their own data centers.

Reflection releases Beam's open weights under Apache 2.0

Reflection published the weights for Beam, its first open model, on Hugging Face on October 16 under the Apache 2.0 license. The model was previewed on October 5 while it was still going through final safety testing, with access limited to a waitlist and the weights promised later in the month. That promise is now kept.

Beam is a mixture-of-experts model, a design that keeps a large pool of parameters but activates only a slice of them for each token. It holds 501 billion total parameters and fires up about 23 billion per token. The context window reaches one million tokens, and Reflection built it for software engineering and long-running agentic work, the kind where a model plans and executes across many steps without a human steering every move.

The size matters less than the license. Apache 2.0 lets a company download Beam, run it behind its own firewall, modify it, and ship products on top without paying Reflection or asking permission. That's a different deal from calling a closed API, and it's the reason a model this large is worth paying attention to.

What ships with the Reflection Beam weights

The release is more than a file. Reflection published the instruction-tuned weights in three precision formats, BF16, FP8, and NVFP4, plus a separate draft model for speculative decoding, which is a trick for generating text faster by guessing ahead and checking the work.

Alongside the weights, Reflection put out a model card with final benchmarks and a safety technical report covering the evaluation approach. The company calls the wider package an "open operating system for agents," with Beam as the kernel and a set of reference integrations and deployment containers layered on top.

The vendor numbers look strong on paper. Reflection reports Beam at 80.9% on SWE-bench Verified, a test of resolving real software issues, and 78.7% on MCP Atlas, which measures tool use through MCP servers. On GPQA Diamond, a graduate-level science reasoning set, it scores 90.5%. Those figures come from Reflection's own evaluations and are repeated in partner launch posts. Independent verification at this scale hasn't caught up yet.

How Reflection trained Beam

Reflection says it pretrained Beam on 23.8 trillion tokens of text. The part that stands out is the reinforcement learning stage, where the model learns by trial and error against a reward signal. The company ran it on 10,500 NVIDIA GB300 GPUs for four weeks and generated more than 100 million rollouts.

A 100-million-rollout training run is how Beam gets better at multi-step problems, where the right answer only appears after a chain of correct moves. Reflection also applied a length penalty during training to discourage padding, its bid to make the model solve tasks while using fewer tokens.

Why should that matter to you? Because cheaper inference is the whole point of an open model. Big models often win benchmarks by thinking longer and spending more. If Beam can match rivals on accuracy while burning fewer tokens, it's cheaper to run, and the cost shows up on the bill every time an agent runs a loop.

Mirror CLI and the model support behind Beam

Reflection open-sourced Mirror CLI, a terminal agent tool trained next to Beam. Mirror pairs a Python frontend for prompts and developer tooling with a Rust core that drives the model. It supports resumable conversations, parallel agents in separate worktrees, project instructions, and MCP connections. Reflection says it built Mirror for internal engineering first, with isolated workspaces and hooks into Slack, Linear, and GitHub, before deciding to release it.

Beam has also been tested against Pi from Earendil, OpenCode from Anomaly, and Hermes from Nous Research on coding and knowledge work, including reading repositories, editing files, and running commands. Reflection opened a public beta of its own API at platform.reflection.ai, which it says is free to start.

The consistent thread is that Reflection wants Beam plugged into the tools developers already use. Owning the weights and the agent tooling together gives the company a full stack to point at, unusual for a launch this size.

What Beam's arrival means for open-source AI

Beam lands at the top end of open-weight models by parameter count, and Apache 2.0 keeps the door open for commercial use. For AI teams weighing whether to build on a closed API, that's a real alternative: a model they can host, tune to their own data, and keep running even if a vendor changes its terms.

The open model has been closing the gap with closed frontier systems for a while now, and a 501-billion-parameter release pushes that trend further. The people who benefit most are teams with data they can't send to a third party: hospitals, banks, defense contractors, and any company with a rule about where its data lives. For them, a strong model they can run behind the firewall removes a blocker that no API can.

The usual catch applies. A 501-billion-parameter mixture-of-experts model needs serious hardware to serve well, which is why the partner list runs long and Dell and HPE are positioned for private deployments. Day-zero support from vLLM, SGLang, and NVIDIA's Dynamo helps, but hosting Beam isn't a laptop project. A team with a single GPU won't be running this model locally.

There's also a governance angle. Apache 2.0 means Reflection can't pull the model back or change the terms after the fact. Once you have the weights, what you do with them is your call. That's the promise open weights always make, and it's why the license matters more than the benchmark table for many buyers.

Two things to watch. One is independent evaluation, since most of the benchmark story is still self-reported. The other is whether the collection of integrations matures past launch-day announcements into maintained tooling. Both will decide whether Beam becomes a default choice or a one-time splash.

How to try Reflection Beam

Beam is available now. The weights are downloadable from Hugging Face, and the launch partners cover cloud, serverless, and on-premises setups. If you want the quickest test, an inference partner that already hosts Beam gets you running without provisioning hardware. If you want control, the downloadable weights and the open Mirror CLI are the path.

Start with a coding task you already know the answer to. Give Beam a real repository and a real issue, and watch how it plans before it edits. That tells you more about the model than any published score. If it holds up, the Apache 2.0 license means you can build on it without asking anyone.