Reflection AI Unveils Beam, a 501B-Parameter Open-Weight Coding Model

On October 5, 2026, Reflection AI published Introducing Beam, its first open-weight model. Beam is a sparse mixture-of-experts with 501 billion total parameters and 23 billion active per token, aimed at coding, reasoning, and agentic workloads. Weights, a technical report, model card, and developer artifacts are promised later in October 2026 under an Apache 2.0 license; an early access waitlist is open while final red-teaming completes.

Training investments Reflection highlights

The post states Beam was pretrained on 23.8 trillion tokens from web and licensed sources. Reflection also describes a large reinforcement-learning phase: more than 100 million rollouts on 10.5K NVIDIA GB300 GPUs over four weeks, with up to 256K-token rollouts and roughly 1.3 billion sandboxes used for training and grading.

Reflection positions that RL scale as among the largest conducted by an open lab to date, with benchmark curves continuing to climb as rollouts increased.

Performance claims—and caveats

Reflection says Beam is competitive with larger open models such as GLM 5.2 and approaching Qwen 3.8-Max on coding and agentic tasks, while noting frontier open models like Kimi K3 remain ahead on raw capability. Beam’s emphasized advantage is inference efficiency: on advanced reasoning benchmarks, Reflection claims scores comparable to GLM-5.2 with roughly 3–4× less inference compute, with broader gaps versus 2T+ models like Qwen 3.8-Max.

Published tables include scores such as 80.1 on Terminal Bench v2.1 (versus 81.0 for GLM 5.2 in the same table) and 80.9 on SWEBench Verified (ahead of Nemotron 3 Ultra’s 70.7 where listed). These are vendor-published numbers; independent replication awaits the weight release.

Reflection also notes efficiency charts exclude prompt prefill, context-dependent attention, and serving overhead—so they represent approximate compute comparisons, not measured dollar costs.

Safety approach

Reflection describes training a separate safety and alignment model from the same checkpoint and merging via distillation, with plans to publish safety evaluation results in the technical report. Beam remains in final red-teaming at announcement time.

The company says it will open-source internal safety evaluations alongside the technical report—unusual transparency for a launch still in red-team. Enterprises evaluating Beam should still run their own policy tests on jailbreaks, tool misuse, and data exfiltration before granting repository access.

Benchmark table highlights

Reflection’s published scorecard spans agentic coding, terminal tasks, tool calling, and general capabilities. Beyond Terminal Bench and SWEBench Verified, the post lists 65.5 on SWE Bench Pro v1 (ahead of GLM 5.2’s 62.1 in the same row), 78.7 on MCP Atlas, and 90.5 on GPQA Diamond. Rows marked NR mean Reflection or competitors have not reported that score publicly.

Readers should treat mixed rows as a signal, not a verdict: Beam trails some models on DeepSWE v1.1 and other agentic suites while leading on select coding verification tasks. The honest takeaway is that open-weight leadership is now multidimensional—terminal agents, repo Q&A, and reasoning efficiency rarely align in one model.

What open-weight competition means for builders

If Apache 2.0 weights ship as promised, teams gain another US-built option for on-prem and VPC deployments—especially relevant alongside NVIDIA’s push for local Spark hardware. The MoE architecture (23B active) is tailored for enterprises that want strong coding agents without activating half-trillion parameters on every token.

Practical next steps: join the waitlist if you need early access, but plan eval harnesses now—SWE-bench variants, internal repos, and agent scenarios that mirror your production tools. Open weights only matter when they beat your incumbent on tasks you actually ship.

RL science details worth noting

Reflection’s post goes deeper on training mechanics than a typical launch blog. The team describes asynchronous policy gradients at scale, policy staleness from long rollouts spanning multiple checkpoints, and algorithmic work to keep learning stable when interactions may be more than a day old. They also cite sourcing roughly one million coding, agentic, and STEM environments for grading—scale that explains why open labs increasingly look like RL factories, not just pretraining shops.

For readers tracking US open-weight strategy, Beam arrives alongside heavy competition from Chinese open models (GLM, Qwen, Kimi families appear in Reflection’s comparison tables). Reflection argues Beam’s wedge is intelligence per flop on Western infrastructure rather than winning every public benchmark row.

Bottom line

Beam is Reflection’s opening move in open weights: massive total parameters, modest active count, heavy RL investment, and a story centered on coding and agentic efficiency rather than sweeping leaderboard domination. October’s weight drop will test whether the efficiency narrative holds once the community runs its own inference bills.