Reflection released Beam 1.0 on October 16, 2026, making its model weights publicly downloadable after an October 5 early-access preview. The release is more than a promise: Reflection’s Hugging Face repository contains the artifacts, with the principal checkpoint listed at approximately 1 TB. 1 2 3
The release offers organizations a route to run and adapt the model rather than rely solely on a hosted API. But the infrastructure requirements are substantial, and external evaluations show a more qualified picture than a blanket claim of frontier leadership or inexpensive deployment. 4 5 6 7
What is available—and what ownership means
Reflection’s model card specifies 501.3 billion total parameters, with 23.4 billion active per token, using a mixture-of-experts architecture. It describes Beam as text-only and predominantly English, advertises a one-million-token context window, and lists instruction-tuned BF16, FP8 and NVFP4 checkpoints. These are the provider’s released specifications, not capabilities independently validated for this draft. 8
The repository includes an Apache 2.0 license. Reflection’s model card expressly permits fine-tuning, quantization, distillation and training other models using Beam’s outputs, while stating that its usage guidance does not restrict the license. 4 8
That makes open-weight under Apache 2.0 the precise description. Downloadable weights and adaptation rights provide operational control, but do not establish full reproducibility of the original training process. Reflection says its training inputs included proprietary licensed material, and the model card still marks a general technical report as forthcoming. 3 2 8
Customization is documented; hardware remains a constraint
The distinction between active parameters and resident model size matters. Hugging Face’s integration guide explains that inactive experts still occupy memory: activating roughly 23 billion parameters does not give Beam the memory footprint of a 23-billion-parameter model. 5
Hugging Face publishes the following reference configurations—not universal minimum requirements: 5
Workload Published reference setup BF16 inference 16 × H200 GPUs FP8 inference 8 × H200 GPUs NVFP4 inference 4 × B300 GPUs LoRA supervised fine-tuning One node with 8 × B300 GPUs Full-parameter GRPO reinforcement learning 96 × H200 GPUs, or 64 × H200 in a colocated configurationThese recipes make customization a concrete option, including adapter-based and full-parameter training. They also place the documented deployments firmly in substantial multi-GPU infrastructure, rather than demonstrating ordinary consumer-hardware accessibility. 5
For users not self-hosting, Reflection also offers a first-party API in public beta and partner access. A separate DFlash2 speculative-decoding drafter is published alongside the main model. 1 9
Performance evidence is useful but uneven
Reflection’s final model card reports BF16 scores of 81.6% on SWE-bench Verified, 80.1% on Terminal-Bench v2.1 and 90.6% on GPQA Diamond, using temperature 0.7 and top_p 0.99. Those are vendor measurements. Hugging Face identifies the principal benchmark table in its integration guide as Reflection-supplied, so that table is not separate performance confirmation. 8 5
External testing adds a different perspective. In its October 16 assessment, Artificial Analysis gave Beam an Intelligence Index score of 30, ranking it second among the US open-weight models it tested and tenth among all open-weight models it tested at publication. It reported roughly 26,000 output tokens per index task, compared with about 52,000 for K2 Horizon, which scored 31. 6
Artificial Analysis also found uneven task performance: 48.5% on AutomationBench-AA and 3.0% on Terminal-Bench 4.0. That latter result is not directly comparable with Reflection’s 80.1% on Terminal-Bench v2.1; different benchmark versions do not establish a like-for-like performance collapse. 6 8
Artificial Analysis provides external measurements, but Reflection also names it among organizations involved in pre-release capability and safety testing. Its results should not be described as evidence of an absence of collaborative or commercial relationships, nor as reproduction of Reflection’s entire benchmark suite. 1 6
Fewer output tokens do not guarantee cheaper tasks
Plotly’s October 17 data-analytics evaluation illustrates the distinction between token efficiency and total task cost. Beam passed 36 of 43 questions, or 84%, matching DeepSeek V4.1 Flash on that exam. Plotly reported a complete-run cost of $8.29 for Beam versus $0.46 for DeepSeek. 7
Beam averaged approximately 15 model calls per question. Only 0.4% of its input tokens were billed as cached, and input charges accounted for $7.79 of the total. Those details matter: the test used a private, synthetic wind-energy dataset, and its small size, domain and hosting conditions limit broader conclusions about relative cost. 7
Reflection’s own claim of a 3–4× inference-compute advantage over GLM-5.2 on selected reasoning comparisons also needs qualification. The company’s approximation uses active parameters and generated tokens while excluding prompt prefill, context-dependent attention operations and serving overhead. It is not a measured end-to-end cost reduction. 2
Taken together, the evaluations support measuring cost per successfully completed task under comparable serving conditions—not inferring it from active-parameter counts or output-token use alone. 6 7
Operational safeguards and unresolved questions
Reflection’s Apache-2.0 Mirror coding harness can read and edit files, execute commands and run parallel tasks in separate repository worktrees. Its repository warns that tools run with the user account’s permissions by default and that manual approval is not a sandbox. 10
Security findings are similarly bounded. Irregular says Reflection evaluated Beam on its private challenges, solving 23 of 40 atomic tasks but none of 10 multistage scenarios. Irregular cautions that results depend on the harness, tools and interaction budget and do not establish real-world attack efficacy. They are not blanket safety certification. 11
Several issues remain open as of the research cutoff:
- Benchmark reconciliation: Hugging Face lists an NVFP4 Terminal-Bench v2.1 score of 78.2%, while Reflection’s model card lists 78.4%. Preview and release results also changed—for example, SWE-bench Pro v2-Hard moved from 77.2% to 76.5%. The supplied research does not explain these differences. 5 8 2
- Production reproducibility: Which exact framework versions, kernels, reasoning settings and hardware configurations reproduce the advertised deployment performance? How do costs compare with equivalent caching, concurrency, context lengths and success criteria? 5 7
- Remaining documentation: The safety technical report is published, but the general technical report remains marked forthcoming. Reflection promised self-distillation recipes after release; this research did not verify their delivery, which is not evidence that they failed to ship. 1 8
Beam’s release therefore establishes a meaningful option for organizations seeking control over model deployment and adaptation. Whether it is the right model to scale remains a workload-specific question involving hardware, reliability and serving economics. 4 5 7
Research cutoff: October 18, 2026. This draft uses the supplied research package only. No hands-on deployment or benchmarking was performed.