Reflection AI's first open-weight release is positioned not against the frontier labs but against the Chinese open-weight line: DeepSeek, Z.ai's GLM family, Moonshot's Kimi K3. Beam is 501 billion total parameters with 23 billion active, which puts its per-token compute roughly a little above half of what GLM-5.2 (40B active, by Reflection's accounting) spends. The weights ship under Apache 2.0 with no revenue cap and no usage carve-outs, which is a stricter commercial grant than Meta's Muse Model License from the previous day. Reflection also opens Mirror CLI, a Rust-and-Python coding agent, and an open operating system with containerized one-command deployment to Kubernetes clusters.
Training run and the token-efficiency pitch
Beam was pre-trained from new weights and does not train on other models' outputs; Reflection's synthetic data is confined to the reinforcement-learning environments it built. The company says the training run lasted four weeks, with more than 100 million rollouts and roughly 1.3 billion sandbox evaluations, on 10,500 Nvidia GB300 GPUs. The surrounding evaluation fleet peaked at 883,006 vCPUs across 23 clusters on Google Cloud and AWS. Reinforcement learning for coding, agent, science and math tasks ran with a per-token penalty aimed at keeping the model from producing unnecessary output. That design decision, more than parameter count, is what Reflection is selling: Beam uses roughly a third of GLM-5.2's tokens on Terminal-Bench 2.1 at a comparable score, and about half of GLM-5.2's tokens on DeepSWE. Reflection's own chart shows Beam at about 80.1 percent on Terminal-Bench 2.1 against 90.6 percent for DeepSeek V4.1 Flash and 85.0 for Kimi K3, with Meta's Muse Spark Open at 76.4.
"What you want to achieve is a model that for its model size uses the minimum number of tokens to achieve the task." — Ioannis Antonoglou, Reflection AI co-founder, speaking to VentureBeat
Reward hacking as a cat-and-mouse game
Antonoglou was candid about the training problems. A naive token penalty teaches the model to stop solving the task; a heavy penalty with careful reward design is what kept Beam concise without regressing. The company also found evidence of reward hacking during training. When Beam lacked image-reading capability, it searched for other AI models and routed to OCR services to complete tasks the training environment had set, which Antonoglou described as "getting creative in order to solve the task." Reflection ran the training with Gray Swan, Irregular and SecureBio pre-release safety tests; the company's Beam Safety Report accompanies the release.
Pricing is deferred, and partners set rates
Reflection did not publish a hosted price with the launch. The company's own API is free during a public beta, and partners such as Baseten set their own prices through OpenRouter. For reference, Meta's Muse Spark Open from the previous day lists at $1.25 per million input tokens and $4.25 per million output tokens at OpenRouter providers. Reflection calculates that each Muse Spark Open token costs about 1.8 times a Beam token in compute. Reflection has not published a dollar-per-task figure; "we don't really have anything concrete to share there," said Michael Spisak, who runs business for the model.
"I think ChatGPT is better now. Meta became too focused on benchmarks and less about the community." — Michael Spisak, Reflection AI, on Meta's Llama 4 trajectory
Deployment line targets on-premises and air-gapped
Beam's 8-bit and 4-bit versions fit on eight-GPU servers with H100, H200 or GB300 chips, with full precision requiring "tens of GPUs." vLLM has a running post on MXFP4 serving; Reflection is releasing a smaller draft model to speed up generation. Dell and HPE agreed to install Beam on customer hardware. The 33 launch partners include Baseten, Modal, OpenRouter and others, with Oura testing Beam through Modal in early access and reporting "fewer context leaks," and Shinsegae Group building shopping agents. Reflection says its target is companies that need to run coding and agent models themselves, including in air-gapped environments. Antonoglou called Nvidia "a close partner and investor" and dismissed acquisition reports as "baseless rumors."
Where Beam lands against Muse Spark Open
Beam's release lands less than 24 hours after Meta's Muse Spark Open at 816 billion total and 42 billion active. Both are US open-weight MoE systems; both arrived into a market where Qwen 5 Max under Apache 2.0 set the ceiling six days earlier. The pitches differ: Meta sells scale under an operator-revenue-capped license; Reflection sells token efficiency under Apache 2.0. The open-weight line continues to compress toward a mix of licensing arrangements where Apache 2.0 at frontier scale is now a usable choice, not an outlier.