Reflection AI's Beam Is Out: 501B Open Weights Under Apache 2.0

Reflection AI made good on its October 5 promise eleven days later: Beam’s weights, technical report, and model card are now public under Apache 2.0, posted on Hugging Face as of October 16. Beam is a sparse Mixture-of-Experts model with 501 billion total parameters and 23 billion active per token, pretrained on 23.8 trillion tokens with effective context stretched to 1 million via midtraining, and tuned for exactly the workloads this site’s readers care about: coding, reasoning, and agents. With Meta’s Muse Spark Open and Muse Glimmer already in the open, the US open-weight field now has a serious third entrant, and Reflection’s benchmark card positions Beam squarely against DeepSeek, Qwen, GLM, and Kimi. The October 5 announcement and the October 16 availability post carry the full specification.

The benchmark card, honestly read

Reflection’s positioning is “intelligence per token,” and the numbers are chosen to support it rather than to claim the frontier. SWE-Bench Verified sits at 80.9, ahead of Inkling (77.6) and Nemotron 3 Ultra (70.7). AIME 2026 hits 97.8 and GPQA Diamond 90.5. The company concedes Kimi K3 still leads on raw capability (Terminal Bench v2.1: Kimi 88.3 and DeepSeek V4.1 Flash 90.6 against Beam’s 80.1) and that Qwen 3.8-Max, a 2T-plus model, outclasses it outright. The claim doing the real work is efficiency: scores comparable to GLM 5.2 on reasoning using 3 to 4 times less inference compute. For enterprise deployments, a workhorse model that is slightly worse per token but four times cheaper changes the total-cost math more than a leaderboard win does.

The RL run is the real engineering story

The pretraining was fast but conventional in scale: under four weeks on 6,144 GB300 NVL72 GPUs at 92.3% goodput. The post-training is where Reflection spent its differentiation: 10.5K GB300 GPUs for four weeks, over 100 million rollouts against roughly a million curated environments and 1.3 billion sandboxes used, with an average of 110,000 concurrent rollouts across 170,000 concurrent sandboxes. The company calls it among the largest RL runs by an open lab, and the infrastructure claims back that up: median weight-distribution time of about 12 seconds across the fleet.

Two algorithmic details stand out. First, fully asynchronous policy gradients that tolerate staleness: the training stayed stable even when some rollouts returned interactions over a day old, computed against policy weights 107 versions behind. That matters because long-horizon agent tasks (the kind Beam targets) cannot be rolled out synchronously without idling the accelerator fleet. Second, a controllable token-efficiency penalty that rewards successful task completion while discouraging wasted tokens; early in training, scores rose while completions got shorter, then lengths grew again as longer reasoning bought more performance. That knob ships to users as a reasoning-effort parameter, the same UX convention OpenAI and Anthropic normalized.

What to watch

Beam’s compute pedigree was never in doubt; we have covered Reflection’s $1B Nebius agreement for GB300 capacity and its SpaceX GB300 cluster deal. What the October 16 release tests is whether open weights plus enterprise-grade agentic benchmarks converts into adoption against incumbents with distribution. Three things are worth tracking: independent reproduction of the SWE-Bench and Terminal Bench numbers once the eval stack is in community hands, how quickly the major serving stacks pick it up (the announcement references ecosystem distribution partners and open-source library integrations without naming them), and whether the asynchronous RL recipe, arguably the most valuable thing in the technical report, gets replicated by other open labs. The 23B-active footprint also puts Beam at the edge of what a multi-GPU workstation can serve, which keeps it out of hobbyist territory but squarely in the self-hosting range that enterprises with data-residency requirements actually buy.