Beam

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.

Teams self-hosting MoE should not plan production environments based on unreleased weights.

Beam's total parameter count is 501B, with 23B parameters activated per token (about 4.6% of the network is activated).

Beam serving takes about 500GB at 8-bit, and about half that at 4-bit.

Reflection claims Beam's reasoning ability is comparable to GLM-5.2, but its reasoning compute is about 1/3 to 1/4 of the latter's.

Benchmark data published by Reflection (not independently verified by third parties): SWE-bench Verified 80.9, Terminal Bench v2.1 80.1, DeepSWE v1.1 44.4.

On Terminal Bench v2.1, Beam's 80.1 trails DeepSeek V4.1 Flash's 90.6; on DeepSWE v1.1, Beam's 44.4 trails GLM-5.3's 61.0 and Kimi K3's 68.0.

Reflection AI claims Beam is on par with the GLM series while requiring 3 to 4 times less inference compute; the company has raised approximately $4.7 billion in cumulative investment from NVIDIA, Sequoia, and others, though the performance figures are self-reported and the weights have not been released.

Early tests by independent benchmarking organization Artificial Analysis show that Beam uses fewer tokens at equivalent intelligence levels.

Beam's full weights are planned to be released under the Apache 2.0 license, but self-deployment is not yet possible, and early access requires joining a waitlist.

Beam's weights, technical report, and model card will be released under the Apache 2.0 license; currently only a waitlist is open and downloading is not yet available.

Beam's pre-training used 6,144 NVIDIA GB300 NVL72 chips, took less than 4 weeks, and consumed 23.8 trillion tokens.

Beam's reinforcement learning phase used 1,500 GB300 chips running continuously for four weeks, generating over 100 million rollouts, with a maximum context length of 256K.

Beam's training prepared approximately 1 million training task environments, opened approximately 1.3 billion sandboxes in total, averaged 110,000 concurrent rollouts, peaked at 170,000 sandboxes, and spanned more than 20 clusters.

During Beam's training, the median time to sync new weights from the training side to the inference cluster was approximately 12 seconds; 71 failures occurred but training never stopped, with a median recovery time of 8 minutes and lost service time accounting for only 0.02%.

On the DeepSWE coding leaderboard, Beam scored 44.4 points, ranking last, below DeepSeek V4.1 Flash's 74.2 points, Kimi K3's approximately 68 points, GLM's approximately 61 points, and Qwen3 Max's approximately 51 points.

On Terminal Bench, Beam scored approximately 81 points, ranking last, below DeepSeek's 90.6, Kimi's 88.3, GLM's 88.2, and Qwen's 86.6.

Reflection officially acknowledged that frontier open-source models such as Kimi K3 still lead Beam in raw capability, and that Beam's advantage lies in efficiency.

Reflection stated that Beam achieves comparable scores to GLM on reasoning benchmarks while consuming three to four times less inference compute, a figure estimated based on activated parameters and generated token counts rather than measured costs.