Reflection AI's Beam: How a 501B MoE Model Learns to Think in Fewer Tokens

Reflection AI released Beam, a 501.3-billion-parameter sparse Mixture-of-Experts model (23.4 billion active parameters per token) under Apache 2.0, trained to use roughly 3–4× fewer inference tokens than comparable open models while matching GLM-5.2 on solution quality. The model was pretrained on 23.8 trillion tokens and post-trained with a controllable length penalty in its RL reward, run across 10,500 NVIDIA GB300 GPUs over four weeks, and ships with a `reasoning_effort` parameter spanning five settings. Reflection also released DFlash2, a 1.93-billion-parameter draft model for speculative decoding that preserves the full model's output distribution.

When most open-weight model releases compete on raw benchmark scores, Reflection AI took a different angle with Beam https://reflection.ai/blog/introducing-beam : instead of asking "how capable can we make this model?", they asked "how much reasoning compute does a correct answer actually require?" The result is a 501-billion-parameter sparse Mixture-of-Experts model that achieves frontier-level coding and agentic performance while using roughly 3–4× fewer inference tokens than comparable open models. Released on October 5, 2026, Beam is available under the Apache 2.0 license with weights on Hugging Face https://huggingface.co/collections/ReflectionAI/beam-67f8e3b2a1c4d5e6f7890abc , making it one of the most capable fully open models for software engineering workloads. Most post-training work focuses on making models smarter at a fixed inference budget. Reflection's approach treats inference compute itself as the thing to optimize. Their key insight is that a model trained to solve problems correctly and concisely is more useful in production than one that reaches the right answer after generating thousands of unnecessary tokens. To operationalize this, they introduced a controllable length penalty into the reinforcement learning reward function. The penalty discourages unnecessary token generation while still rewarding successful task completion. The result is a model that learns to match its reasoning depth to the actual complexity of the task at hand. Reflection estimates inference efficiency using a simple formula: generation compute ≈ 2 × active parameters × average tokens generated per attempt . By this measure, Beam achieves comparable solution quality to GLM-5.2 while using 3–4× less generation compute — not by being less capable, but by being less verbose. Beam's architecture is a sparse MoE with 501.3 billion total parameters and 23.4 billion active parameters per token. That active-parameter count is what matters for inference cost: each forward pass only activates a small fraction of the full model. Several design choices prioritize numerical stability, which matters especially for downstream RL training: The model was pretrained on 23.8 trillion high-quality tokens, then put through one of the largest RL runs conducted by an open-weight lab: over 100 million rollouts generated across 10,500 NVIDIA GB300 GPUs over four weeks, using asynchronous policy gradients https://venturebeat.com/technology/reflections-beam-aims-to-undercut-chinese-open-models-with-fewer-tokens-not-cheaper-tokens to handle the staleness that comes with training at that scale. The RL training unfolded in two distinct phases, which Reflection describes as a moving Pareto frontier between capability and token usage: Phase 1 — Contraction: Early in RL, the model prioritized efficiency. It learned to solve the same problems it could already solve, but with shorter completions. Performance held steady while token counts dropped. Phase 2 — Expansion: As the model acquired more advanced agentic capabilities — multi-step tool use, web browsing, complex code navigation — completion lengths grew again. But this growth was purposeful: the additional tokens were doing real reasoning work, not padding. The net effect is a model that has internalized when to think briefly and when to think deeply. Users can also control this explicitly via a reasoning effort parameter with five settings from low to max , letting developers tune the compute-quality trade-off for their specific workload. To address wall-clock latency separately from token count, Reflection released a companion model: DFlash2 https://reflection.ai/beam-model-card , a 1.93-billion-parameter draft model designed for speculative decoding. DFlash2 generates candidate token sequences ahead of the main Beam model, which then verifies them in batches using rejection sampling. This preserves the exact output distribution of the full model while accelerating generation by 1.72× to 3.03× depending on task and hardware configuration. The NVFP4 version of Beam with DFlash2 shows the largest speedups, making it practical to run on eight-GPU configurations. On software engineering benchmarks, Beam posts competitive numbers: These numbers are notable — independent analysis from Marktechpost https://www.marktechpost.com/2026/10/05/reflection-ai-introduces-beam-a-501b-open-weight-moe-model-with-23b-active-parameters-for-coding-and-agentic-workloads/ confirms the efficiency claims — because they come from a model that is simultaneously optimizing for fewer tokens. The usual assumption is that token efficiency and task performance trade off against each other; Beam's results suggest that with the right RL setup, you can improve both simultaneously — at least up to a point. Beam is available in BF16, FP8, and NVFP4 formats, with deployment support via Amazon SageMaker JumpStart https://aws.amazon.com/blogs/machine-learning/introducing-reflection-beam-an-open-weight-model-for-agentic-coding-on-aws/ and Kubernetes clusters running vLLM or SGLang. The Apache 2.0 license explicitly permits fine-tuning, quantization, distillation, and model merging, which is a meaningful contrast to some open-weight releases that restrict downstream use. The 1-million-token context window makes it viable for codebase-wide tasks — full repository analysis, large-scale refactoring, or long-horizon agentic workflows that need to hold a lot of state in context. A detailed technical breakdown from Xenospectrum https://xenospectrum.com/en/reflection-beam-open-model-efficiency/ covers the serving infrastructure in more depth, including the fused kernels Reflection developed to improve throughput on multi-GPU setups. The practical implication of Beam's approach is cost. If a model uses 3–4× fewer tokens to reach the same answer, API costs drop proportionally for generation-heavy workloads. For teams running large-scale agentic pipelines — automated code review, continuous integration agents, research assistants — that multiplier compounds quickly. The reasoning effort parameter also gives operators a lever that most models don't expose cleanly. Rather than choosing between a fast cheap model and a slow expensive one, you can run the same model at different effort levels and tune the trade-off per task type. The open licensing and multi-format release BF16/FP8/NVFP4 lower the barrier to self-hosting, which matters for teams with data-residency requirements or those who want to fine-tune on proprietary codebases. Beam represents a coherent thesis about where open-weight model development should go: rather than simply scaling parameters or context windows, optimize for how much compute a correct answer actually requires. The combination of a length-penalized RL objective, a stable MoE architecture, and a speculative decoding companion model gives practitioners a tool that is both capable and economical to run. Whether this approach generalizes beyond coding and agentic tasks remains to be seen, but as a proof of concept for inference-compute efficiency as a first-class training objective, it makes a strong case.